From 2b044fda29389b25dfc4f4f004a39569da307f21 Mon Sep 17 00:00:00 2001 From: dengwenjie <775747758@qq.com> Date: Sun, 31 Aug 2025 18:41:27 +0800 Subject: [PATCH] =?UTF-8?q?=E4=B8=B4=E6=97=B6=E6=8F=90=E4=BA=A4?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../common/utils/DJLCommonUtils.java | 14 + .../common/utils/LetterBoxUtils.java | 67 ++- .../cn/smartjavaai/common/utils/NMSUtils.java | 69 ++- examples/face-example/pom.xml | 14 +- .../main/java/smartai/examples/face/Test.java | 22 +- .../face/expression/ExpressionRecDemo.java | 18 +- .../examples/face/facedet/FaceDetDemo.java | 102 ++-- .../examples/face/facerec/FaceRecDemo.java | 78 ++- .../face/liveness/LivenessDetDemo.java | 18 +- examples/objectdetection-example/pom.xml | 10 +- examples/ocr-examples/pom.xml | 10 +- examples/translation-example/pom.xml | 10 +- .../face/enums/FaceDetModelEnum.java | 14 +- .../face/factory/FaceDetModelFactory.java | 5 +- .../model/facedect/MtcnnFaceDetModel.java | 88 +++- .../criterial/FaceDetCriteriaFactory.java | 33 +- .../face/model/facedect/mtcnn/ONetModel.java | 5 +- .../face/model/facedect/mtcnn/PNetModel.java | 103 ++-- .../face/model/facedect/mtcnn/RNetModel.java | 2 +- .../translator/FaceDetectionTranslator.java | 50 +- .../face/translator/SCRFDFaceTranslator.java | 242 --------- .../face/translator/YoloV5FaceTranslator.java | 154 +++--- .../face/translator/YoloV8FaceTranslator.java | 476 ------------------ .../cn/smartjavaai/face/utils/FaceUtils.java | 7 +- face/src/test/java/Test.java | 11 +- 25 files changed, 617 insertions(+), 1005 deletions(-) delete mode 100644 face/src/main/java/cn/smartjavaai/face/translator/SCRFDFaceTranslator.java delete mode 100644 face/src/main/java/cn/smartjavaai/face/translator/YoloV8FaceTranslator.java diff --git a/common/src/main/java/cn/smartjavaai/common/utils/DJLCommonUtils.java b/common/src/main/java/cn/smartjavaai/common/utils/DJLCommonUtils.java index 0f1ccb1..61caac2 100644 --- a/common/src/main/java/cn/smartjavaai/common/utils/DJLCommonUtils.java +++ b/common/src/main/java/cn/smartjavaai/common/utils/DJLCommonUtils.java @@ -1,7 +1,11 @@ package cn.smartjavaai.common.utils; +import ai.djl.ndarray.NDArray; +import cn.smartjavaai.common.entity.R; + import java.nio.file.Files; import java.nio.file.Path; +import java.util.Objects; /** * @author dwj @@ -29,4 +33,14 @@ public class DJLCommonUtils { return Files.exists(servingFile); } + /** + * 判断 NDArray 是否为空 + * @param ndArray + * @return + */ + public static boolean isNDArrayEmpty(NDArray ndArray){ + return Objects.isNull(ndArray) || ndArray.size() == 0; + } + + } diff --git a/common/src/main/java/cn/smartjavaai/common/utils/LetterBoxUtils.java b/common/src/main/java/cn/smartjavaai/common/utils/LetterBoxUtils.java index c2eca79..89eaa93 100644 --- a/common/src/main/java/cn/smartjavaai/common/utils/LetterBoxUtils.java +++ b/common/src/main/java/cn/smartjavaai/common/utils/LetterBoxUtils.java @@ -1,5 +1,7 @@ package cn.smartjavaai.common.utils; +import ai.djl.modality.cv.output.Landmark; +import ai.djl.modality.cv.output.Point; import ai.djl.modality.cv.output.Rectangle; import ai.djl.modality.cv.util.NDImageUtils; import ai.djl.ndarray.NDArray; @@ -8,7 +10,9 @@ import ai.djl.ndarray.index.NDIndex; import ai.djl.ndarray.types.DataType; import ai.djl.ndarray.types.Shape; +import java.util.ArrayList; import java.util.Arrays; +import java.util.List; /** * 按比例缩放，剩余空间用指定颜色填充 @@ -61,7 +65,7 @@ public class LetterBoxUtils { // NDArray paddingImg = manager // .full(new Shape(targetW, targetH, 3), padColor, DataType.UINT8); - NDArray paddingImg = manager.zeros(new Shape(targetW, targetH, 3), DataType.FLOAT32); + NDArray paddingImg = manager.zeros(new Shape(targetH, targetW, 3), DataType.FLOAT32); paddingImg = paddingImg.add(114); int padW = targetW - newW; @@ -145,4 +149,65 @@ public class LetterBoxUtils { return new Rectangle(x1, y1, boxW, boxH); } + /** + * 恢复缩放后的 box(左上角坐标) + * @param landmark + * @param scale + * @param origImageWidth + * @param origImageHeight + */ + public static Landmark restoreBox(Landmark landmark, float scale, int origImageWidth, int origImageHeight, int inputWidth, int inputHeight, boolean isNormalized){ + double x = 0; + double y = 0; + double width = 0; + double height = 0; + if(isNormalized){ + x = landmark.getX() * inputWidth; + y = landmark.getY() * inputHeight; + width = landmark.getWidth() * inputWidth; + height = landmark.getHeight() * inputHeight; + }else{ + x = landmark.getX(); + y = landmark.getY(); + width = landmark.getWidth(); + height = landmark.getHeight(); + } + double paddingWidth = (inputWidth - origImageWidth * scale) / 2; + double paddingHeight = (inputHeight - origImageHeight * scale) / 2; + + // 去掉 padding + double x_noPad = x - paddingWidth; + double y_noPad = y - paddingHeight; + + //模型输出就是原图坐标 + double x1 = x_noPad / scale / origImageWidth; + double y1 = y_noPad / scale / origImageHeight; + double boxW = width / scale / origImageWidth ; + double boxH = height / scale / origImageHeight; + + List points = new ArrayList<>(); + // 要求关键点未归一化 + landmark.getPath().forEach(point -> { + double pointX = (point.getX() - paddingWidth) / scale; + double pointY = (point.getY() - paddingHeight) / scale; + points.add(new Point(pointX, pointY)); + }); + return new Landmark(x1, y1, boxW, boxH, points); + } + + /** + * 获取缩放后的图片大小 + * @param origW 原始图片宽度 + * @param origH 原始图片高度 + * @param targetWidth 目标图片宽度 + * @param targetHeight 目标图片高度 + * @return + */ + public static int[] getResizeSize(int origW, int origH, int targetWidth, int targetHeight){ + float r = Math.min(targetWidth / (float) origW, targetHeight / (float) origH); + int newW = Math.round(origW * r); + int newH = Math.round(origH * r); + return new int[]{newW, newH}; + } + } diff --git a/common/src/main/java/cn/smartjavaai/common/utils/NMSUtils.java b/common/src/main/java/cn/smartjavaai/common/utils/NMSUtils.java index 390f892..35bbd2a 100644 --- a/common/src/main/java/cn/smartjavaai/common/utils/NMSUtils.java +++ b/common/src/main/java/cn/smartjavaai/common/utils/NMSUtils.java @@ -81,16 +81,23 @@ public class NMSUtils { * */ public static NDArray batchedNms(NDArray boxes, NDArray scores, NDArray idxs, float iouThreshold, NDManager manager) { + +// System.out.println("---------------boxes:" + Arrays.toString(boxes.toFloatArray())); + List keepList = new ArrayList<>(); + // 获取唯一 batch id NDArray uniqueIdxs = idxs.unique().get(0); + for (long batchId : uniqueIdxs.toLongArray()) { // 找出当前 batch 的框 NDArray mask = idxs.eq(batchId); NDArray batchBoxes = boxes.get(mask); NDArray batchScores = scores.get(mask); + // 执行单 batch NMS - int[] keepIndices = nms(batchBoxes, batchScores, iouThreshold); + int[] keepIndices = mtcnnNms(batchBoxes, batchScores, iouThreshold); + if (keepIndices.length > 0) { // 将局部索引映射回全局索引 NDArray globalIndices = manager.arange(boxes.getShape().get(0)) @@ -101,10 +108,70 @@ public class NMSUtils { keepList.add(globalIndices); } } + if (keepList.isEmpty()) { return manager.create(new long[0]); } return NDArrays.concat(new NDList(keepList)); } + + public static int[] mtcnnNms(NDArray boxes, NDArray scores, float iouThreshold) { + if (boxes.isEmpty()) { + return new int[0]; + } + + NDArray x1 = boxes.get(":, 0"); + NDArray y1 = boxes.get(":, 1"); + NDArray x2 = boxes.get(":, 2"); + NDArray y2 = boxes.get(":, 3"); + + // 面积 + NDArray areas = x2.sub(x1).add(1).mul(y2.sub(y1).add(1)); + + // scores 降序索引 + NDArray order = scores.argSort(); + //System.out.println("order：" + order.getShape()); + //System.out.println("order：" + Arrays.toString(order.toLongArray())); + + List keep = new ArrayList<>(); + + while (order.size() > 0) { + int i = (int) order.getLong(-1); + keep.add(i); + + if (order.size() == 1) break; // 没框了就退出 + + // 剩余框 + NDArray idx = order.get("0:-1"); + + NDArray xx1 = x1.get(i).maximum(x1.get(idx)); + NDArray yy1 = y1.get(i).maximum(y1.get(idx)); + NDArray xx2 = x2.get(i).minimum(x2.get(idx)); + NDArray yy2 = y2.get(i).minimum(y2.get(idx)); + + NDArray w = xx2.sub(xx1).add(1).maximum(0); + NDArray h = yy2.sub(yy1).add(1).maximum(0); + NDArray inter = w.mul(h); + + NDArray union = areas.get(i).minimum(areas.get(idx)); + NDArray iou = inter.div(union); + +// System.out.println("Max IoU: " + iou.max().getFloat()); +// System.out.println("Min IoU: " + iou.min().getFloat()); +// System.out.println("Mean IoU: " + iou.mean().getFloat()); + +// System.out.println("Before: " + order.size()); + // 保留 IoU <= 阈值的框 + NDArray mask = iou.lte(iouThreshold); +// System.out.println("Mask size: " + mask.size() + " True count: " + mask.sum()); + + // 更新 order + order = idx.get(mask); +// System.out.println("After: " + order.size()); + } + + return keep.stream().mapToInt(Integer::intValue).toArray(); + } + } diff --git a/examples/face-example/pom.xml b/examples/face-example/pom.xml index d10c33d..5562054 100644 --- a/examples/face-example/pom.xml +++ b/examples/face-example/pom.xml @@ -34,7 +34,7 @@ cn.smartjavaai - smartjavaai-bom + bom ${smartjavaai.version} pom @@ -94,14 +94,14 @@ cn.smartjavaai - smartjavaai-face + face ai.djl.pytorch pytorch-jni - 2.5.1-0.32.0 + 2.7.1-0.34.0 runtime @@ -138,7 +138,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.windows-x86_64} - 2.5.1 + 2.7.1 runtime @@ -176,7 +176,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.linux-x86_64} - 2.5.1 + 2.7.1 runtime @@ -213,7 +213,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.osx-aarch64} - 2.5.1 + 2.7.1 runtime @@ -260,7 +260,7 @@ pytorch-native-cpu linux-aarch64 runtime - 2.5.1 + 2.7.1 diff --git a/examples/face-example/src/main/java/smartai/examples/face/Test.java b/examples/face-example/src/main/java/smartai/examples/face/Test.java index cbcbcd7..67b3870 100644 --- a/examples/face-example/src/main/java/smartai/examples/face/Test.java +++ b/examples/face-example/src/main/java/smartai/examples/face/Test.java @@ -84,16 +84,16 @@ public class Test { log.info("user ./gradlew listModel --args='-a' to show artifact detail"); log.info("============================================================"); } - Map> models = ModelZoo.listModels(); - for (Map.Entry> entry : models.entrySet()) { - String appName = entry.getKey().toString(); - for (Artifact artifact : entry.getValue()) { - if (withArtifacts) { - log.info("{} djl://{}", appName, artifact); - } else { - log.info("{} {}", appName, artifact); - } - } - } +// Map> models = ModelZoo.listModels(); +// for (Map.Entry> entry : models.entrySet()) { +// String appName = entry.getKey().toString(); +// for (Artifact artifact : entry.getValue()) { +// if (withArtifacts) { +// log.info("{} djl://{}", appName, artifact); +// } else { +// log.info("{} {}", appName, artifact); +// } +// } +// } } } diff --git a/examples/face-example/src/main/java/smartai/examples/face/expression/ExpressionRecDemo.java b/examples/face-example/src/main/java/smartai/examples/face/expression/ExpressionRecDemo.java index 8e351db..b37e62e 100644 --- a/examples/face-example/src/main/java/smartai/examples/face/expression/ExpressionRecDemo.java +++ b/examples/face-example/src/main/java/smartai/examples/face/expression/ExpressionRecDemo.java @@ -63,15 +63,21 @@ public class ExpressionRecDemo { } /** - * 获取人脸检测模型 + * 获取人脸检测模型（均衡模型） + * 均衡模型：兼顾速度和精度 + * 注意事项：SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath * @return */ public FaceDetModel getFaceDetModel(){ FaceDetConfig config = new FaceDetConfig(); - config.setModelEnum(FaceDetModelEnum.RETINA_FACE);//人脸检测模型 - config.setConfidenceThreshold(FaceDetectConstant.DEFAULT_CONFIDENCE_THRESHOLD);//只返回相似度大于该值的人脸 - config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD);//用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 - config.setDevice(device); + //人脸检测模型，SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath + config.setModelEnum(FaceDetModelEnum.MTCNN); + //下载模型并替换本地路径，下载地址：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234 + config.setModelPath("/Users/wenjie/Documents/develop/face_model"); + //只返回相似度大于该值的人脸,需要根据实际情况调整，分值越大越严格容易漏检，分值越小越宽松容易误识别 + config.setConfidenceThreshold(0.5f); + //用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 + config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD); return FaceDetModelFactory.getInstance().getModel(config); } @@ -83,7 +89,7 @@ public class ExpressionRecDemo { public ExpressionModel getExpressionModel(){ FaceExpressionConfig config = new FaceExpressionConfig(); config.setModelEnum(ExpressionModelEnum.FrEmotion); - config.setModelPath("/Users/xxx/Documents/develop/model/emotion/fr_expression.onnx"); + config.setModelPath("/Users/wenjie/Documents/develop/model/emotion/fr_expression.onnx"); config.setDevice(device); config.setAlign(true); config.setDetectModel(getFaceDetModel()); diff --git a/examples/face-example/src/main/java/smartai/examples/face/facedet/FaceDetDemo.java b/examples/face-example/src/main/java/smartai/examples/face/facedet/FaceDetDemo.java index 52dc205..99f7192 100644 --- a/examples/face-example/src/main/java/smartai/examples/face/facedet/FaceDetDemo.java +++ b/examples/face-example/src/main/java/smartai/examples/face/facedet/FaceDetDemo.java @@ -58,18 +58,58 @@ public class FaceDetDemo { /** - * 获取人脸检测模型 - * 注意事项：高精度模型，速度较慢 + * 获取人脸检测模型（均衡模型） + * 均衡模型：兼顾速度和精度 + * 注意事项：SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath * @return */ public FaceDetModel getFaceDetModel(){ FaceDetConfig config = new FaceDetConfig(); - //高精度模型，速度慢 - config.setModelEnum(FaceDetModelEnum.RETINA_FACE);//人脸检测模型 + //人脸检测模型，SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath + config.setModelEnum(FaceDetModelEnum.MTCNN); + //下载模型并替换本地路径，下载地址：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234 + config.setModelPath("/Users/wenjie/Documents/develop/face_model"); + //只返回相似度大于该值的人脸,需要根据实际情况调整，分值越大越严格容易漏检，分值越小越宽松容易误识别 + config.setConfidenceThreshold(0.5f); + //用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 + config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD); + return FaceDetModelFactory.getInstance().getModel(config); + } + + + /** + * 获取人脸检测模型（高精度模型） + * 注意事项：高精度模型，识别准确度高，速度慢 + * @return + */ + public FaceDetModel getProFaceDetModel(){ + FaceDetConfig config = new FaceDetConfig(); + //人脸检测模型，SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath + config.setModelEnum(FaceDetModelEnum.RETINA_FACE); //下载模型并替换本地路径，下载地址：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234 config.setModelPath("/Users/wenjie/Documents/develop/face_model/retinaface.pt"); - config.setConfidenceThreshold(FaceDetectConstant.DEFAULT_CONFIDENCE_THRESHOLD);//只返回相似度大于该值的人脸 - config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD);//用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 + //只返回相似度大于该值的人脸,需要根据实际情况调整，分值越大越严格容易漏检，分值越小越宽松容易误识别 + config.setConfidenceThreshold(0.5f); + //用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 + config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD); + return FaceDetModelFactory.getInstance().getModel(config); + } + + /** + * 获取人脸检测模型（极速模型） + * 注意事项：极速模型，识别准确度低，速度快 + * @return + */ + public FaceDetModel getFastFaceDetModel(){ + FaceDetConfig config = new FaceDetConfig(); + //人脸检测模型，SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath + config.setModelEnum(FaceDetModelEnum.YOLOV5_FACE_320); + //下载模型并替换本地路径，下载地址：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234 + config.setModelPath("/Users/wenjie/Documents/develop/face_model/yolo-face/yolov5face-n-0.5-320x320.onnx"); + //只返回相似度大于该值的人脸,需要根据实际情况调整，分值越大越严格容易漏检，分值越小越宽松容易误识别 + config.setConfidenceThreshold(0.5f); + //用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 + config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD); return FaceDetModelFactory.getInstance().getModel(config); } @@ -84,47 +124,21 @@ public class FaceDetDemo { config.setModelEnum(FaceDetModelEnum.SEETA_FACE6_MODEL); //指定模型路径：请根据实际情况替换为本地模型文件的绝对路径（下载地址：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234） config.setModelPath("C:/Users/Administrator/Downloads/sf3.0_models/sf3.0_models"); + config.setConfidenceThreshold(0.9); return FaceDetModelFactory.getInstance().getModel(config); } + /** - * 人脸检测(默认配置) - * 使用默认模型参数检测，默认模型：retinaface，需联网，会自动下载模型 - * 图片参数：图片路径 + * 人脸检测 + * 注意事项： + * 1、此用例使用均衡模型，可以切换高精度模型或极速模型 */ @Test public void testFaceDetect(){ try { FaceDetModel faceModel = getFaceDetModel(); R detectedResult = faceModel.detect(imgPath); -// if(detectedResult.isSuccess()){ -// log.info("人脸检测结果：{}", JSONObject.toJSONString(detectedResult.getData())); -// }else{ -// log.info("人脸检测失败：{}", detectedResult.getMessage()); -// } - - long start = System.currentTimeMillis(); - R detectedResult2 = faceModel.detect("/Users/wenjie/Downloads/facetest/surprise.png"); - log.info("耗时：{}", System.currentTimeMillis() - start); - if(detectedResult2.isSuccess()){ - log.info("人脸检测结果：{}", JSONObject.toJSONString(detectedResult2.getData())); - }else{ - log.info("人脸检测失败：{}", detectedResult2.getMessage()); - } - } catch (Exception e) { - e.printStackTrace(); - } - } - - /** - * 人脸检测(自定义模型参数) - * 图片参数：图片路径 - */ - @Test - public void testFaceDetectCustomConfig(){ - try { - FaceDetModel faceModel = getFaceDetModel(); - R detectedResult = faceModel.detect(imgPath); if(detectedResult.isSuccess()){ log.info("人脸检测结果：{}", JSONObject.toJSONString(detectedResult.getData())); }else{ @@ -181,10 +195,12 @@ public class FaceDetDemo { public void testDetectFaceGPU(){ try { FaceDetConfig config = new FaceDetConfig(); - //高精度模型，速度慢 - config.setModelEnum(FaceDetModelEnum.RETINA_FACE);//人脸检测模型 + //人脸检测模型，SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath + config.setModelEnum(FaceDetModelEnum.MTCNN); //下载模型并替换本地路径，下载地址：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234 - config.setModelPath("/Users/xxx/Documents/develop/model/retinaface.pt"); + config.setModelPath("/Users/wenjie/Documents/develop/face_model"); + //只返回相似度大于该值的人脸,需要根据实际情况调整，分值越大越严格容易漏检，分值越小越宽松容易误识别 + config.setConfidenceThreshold(0.5f); config.setDevice(DeviceEnum.GPU); FaceDetModel faceModel = FaceDetModelFactory.getInstance().getModel(config); R detectedResult = faceModel.detect(imgPath); @@ -200,7 +216,7 @@ public class FaceDetDemo { /** * 人脸检测(Seetaface6) - * 图片参数：图片路径 + * 注意事项：不支持macos */ @Test public void testFaceDetectSeetaface6(){ @@ -220,12 +236,12 @@ public class FaceDetDemo { /** * 摄像头人脸检测 - * 注意事项：如果视频比较卡，可以使用轻量的人脸检测模型 + * 注意事项：实时检测，需要使用极速模型 */ @Test public void testDetectCamera(){ try { - FaceDetModel faceModel = getSeetaface6DetModel(); + FaceDetModel faceModel = getFastFaceDetModel(); OpenCV.loadShared(); VideoCapture capture = new VideoCapture(0); if (!capture.isOpened()) { diff --git a/examples/face-example/src/main/java/smartai/examples/face/facerec/FaceRecDemo.java b/examples/face-example/src/main/java/smartai/examples/face/facerec/FaceRecDemo.java index 38f0495..cc9d3d2 100644 --- a/examples/face-example/src/main/java/smartai/examples/face/facerec/FaceRecDemo.java +++ b/examples/face-example/src/main/java/smartai/examples/face/facerec/FaceRecDemo.java @@ -53,39 +53,65 @@ public class FaceRecDemo { /** - * 获取人脸检测模型（高精度，速度慢） - * 追求准确度可以使用 - * 也可以使用其他模型，具体其他模型参数可以查看文档：http://doc.smartjavaai.cn/face.html + * 获取人脸检测模型（均衡模型） + * 均衡模型：兼顾速度和精度 + * 注意事项：SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath * @return */ - public FaceDetModel getHighAccuracyDetModel(){ + public FaceDetModel getFaceDetModel(){ FaceDetConfig config = new FaceDetConfig(); - //高精度模型，速度慢 - config.setModelEnum(FaceDetModelEnum.RETINA_FACE);//人脸检测模型 + //人脸检测模型，SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath + config.setModelEnum(FaceDetModelEnum.MTCNN); //下载模型并替换本地路径，下载地址：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234 -// config.setModelPath("/Users/wenjie/Documents/develop/model/retinaface.pt"); - config.setConfidenceThreshold(FaceDetectConstant.DEFAULT_CONFIDENCE_THRESHOLD);//只返回相似度大于该值的人脸 - config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD);//用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 - config.setDevice(device); + config.setModelPath("/Users/wenjie/Documents/develop/face_model"); + //只返回相似度大于该值的人脸,需要根据实际情况调整，分值越大越严格容易漏检，分值越小越宽松容易误识别 + config.setConfidenceThreshold(0.5f); + //用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 + config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD); return FaceDetModelFactory.getInstance().getModel(config); } /** - * 获取人脸检测模型（高速模型，精度一般） - * 追求速度可以使用 - * 也可以使用其他模型，具体其他模型参数可以查看文档：http://doc.smartjavaai.cn/face.html + * 获取人脸检测模型（高精度模型） + * 注意事项： + * 1、高精度模型，识别准确度高，速度慢 + * 2、具体其他模型参数可以查看文档：http://doc.smartjavaai.cn/face.html * @return */ - public FaceDetModel getHighSpeedDetModel(){ + public FaceDetModel getProFaceDetModel(){ FaceDetConfig config = new FaceDetConfig(); - //高速模型，速度快，精度一般 - config.setModelEnum(FaceDetModelEnum.SEETA_FACE6_MODEL); + //人脸检测模型，SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath + config.setModelEnum(FaceDetModelEnum.RETINA_FACE); //下载模型并替换本地路径，下载地址：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234 - config.setModelPath("/Users/xxx/Documents/develop/model/sf3.0_models"); - config.setDevice(device); + config.setModelPath("/Users/wenjie/Documents/develop/face_model/retinaface.pt"); + //只返回相似度大于该值的人脸,需要根据实际情况调整，分值越大越严格容易漏检，分值越小越宽松容易误识别 + config.setConfidenceThreshold(0.5f); + //用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 + config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD); return FaceDetModelFactory.getInstance().getModel(config); } + /** + * 获取人脸检测模型（极速模型） + * 注意事项： + * 1、极速模型，识别准确度低，速度快 + * 2、具体其他模型参数可以查看文档：http://doc.smartjavaai.cn/face.html + * @return + */ + public FaceDetModel getFastFaceDetModel(){ + FaceDetConfig config = new FaceDetConfig(); + //人脸检测模型，SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath + config.setModelEnum(FaceDetModelEnum.YOLOV5_FACE_320); + //下载模型并替换本地路径，下载地址：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234 + config.setModelPath("/Users/wenjie/Documents/develop/face_model/yolo-face/yolov5face-n-0.5-320x320.onnx"); + //只返回相似度大于该值的人脸,需要根据实际情况调整，分值越大越严格容易漏检，分值越小越宽松容易误识别 + config.setConfidenceThreshold(0.5f); + //用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 + config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD); + return FaceDetModelFactory.getInstance().getModel(config); + } + + /** * 获取人脸识别模型（高精度，速度慢） * 追求准确度可以使用 @@ -97,14 +123,14 @@ public class FaceRecDemo { //高精度模型，速度慢 config.setModelEnum(FaceRecModelEnum.ELASTIC_FACE_MODEL); //模型路径，请下载模型并替换为本地路径：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234 - config.setModelPath("/Users/xxx/Documents/develop/model/elasticface.pt"); + config.setModelPath("/Users/wenjie/Documents/develop/model/elasticface.pt"); //裁剪人脸：如果图片已经是裁剪过的，则请将此参数设置为false config.setCropFace(true); //开启人脸对齐：适用于人脸不正的场景，开启将提升人脸特征准确度，关闭可以提升性能 config.setAlign(true); config.setDevice(device); //指定人脸检测模型 - config.setDetectModel(getHighAccuracyDetModel()); + config.setDetectModel(getProFaceDetModel()); return FaceRecModelFactory.getInstance().getModel(config); } @@ -126,7 +152,7 @@ public class FaceRecDemo { config.setAlign(false); config.setDevice(device); //指定人脸检测模型 - config.setDetectModel(getHighSpeedDetModel()); + config.setDetectModel(getFastFaceDetModel()); return FaceRecModelFactory.getInstance().getModel(config); } @@ -137,14 +163,14 @@ public class FaceRecDemo { public FaceRecModel getFaceRecModelWithDbConfig(){ FaceRecConfig config = new FaceRecConfig(); //高精度模型，速度慢,追求速度请更换高速模型，具体其他模型参数可以查看文档：http://doc.smartjavaai.cn/face.html - config.setModelEnum(FaceRecModelEnum.ELASTIC_FACE_MODEL);//人脸检测模型 + config.setModelEnum(FaceRecModelEnum.ELASTIC_FACE_MODEL);//人脸识别模型 config.setModelPath("/Users/xxx/Documents/develop/model/elasticface.pt"); //裁剪人脸：如果图片已经是裁剪过的，则请将此参数设置为false config.setCropFace(true); //开启人脸对齐：适用于人脸不正的场景，开启将提升人脸特征准确度，关闭可以提升性能 config.setAlign(true); - //指定人脸检测模型，高精度模型，速度慢,追求速度请更换高速模型getHighSpeedDetModel - config.setDetectModel(getHighAccuracyDetModel()); + //指定人脸检测模型，可切换人脸检测模型（极速：getFastFaceDetModel，高精度：getProFaceDetModel），具体其他模型参数可以查看文档：http://doc.smartjavaai.cn/face.html + config.setDetectModel(getFaceDetModel()); config.setDevice(device); //初始化向量数据库：Milvus数据库配置 @@ -175,8 +201,8 @@ public class FaceRecDemo { config.setCropFace(true); //开启人脸对齐：适用于人脸不正的场景，开启将提升人脸特征准确度，关闭可以提升性能 config.setAlign(true); - //指定人脸检测模型，高精度模型，速度慢,追求速度请更换高速模型getHighSpeedDetModel - config.setDetectModel(getHighAccuracyDetModel()); + //指定人脸检测模型，可切换人脸检测模型（极速：getFastFaceDetModel，高精度：getProFaceDetModel），具体其他模型参数可以查看文档：http://doc.smartjavaai.cn/face.html + config.setDetectModel(getFaceDetModel()); config.setDevice(device); //初始化SQLite数据库 diff --git a/examples/face-example/src/main/java/smartai/examples/face/liveness/LivenessDetDemo.java b/examples/face-example/src/main/java/smartai/examples/face/liveness/LivenessDetDemo.java index 6f3dfb8..b1dedd8 100644 --- a/examples/face-example/src/main/java/smartai/examples/face/liveness/LivenessDetDemo.java +++ b/examples/face-example/src/main/java/smartai/examples/face/liveness/LivenessDetDemo.java @@ -80,7 +80,7 @@ public class LivenessDetDemo { config.setModelEnum(LivenessModelEnum.IIC_FL_MODEL); config.setDevice(device); //需替换为实际模型存储路径 - config.setModelPath("/Users/xxx/Documents/develop/model/anti/IIC_Fl.onnx"); + config.setModelPath("/Users/wenjie/Documents/develop/model/anti/IIC_Fl.onnx"); //人脸活体阈值,可选,默认0.8，超过阈值则认为是真人，低于阈值是非活体 config.setRealityThreshold(LivenessConstant.DEFAULT_REALITY_THRESHOLD); /*视频检测帧数，可选，默认10，输出帧数超过这个number之后，就可以输出识别结果。 @@ -122,15 +122,21 @@ public class LivenessDetDemo { /** - * 获取人脸检测模型 + * 获取人脸检测模型（均衡模型） + * 均衡模型：兼顾速度和精度 + * 注意事项：SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath * @return */ public FaceDetModel getFaceDetModel(){ FaceDetConfig config = new FaceDetConfig(); - config.setModelEnum(FaceDetModelEnum.RETINA_FACE);//人脸检测模型 - config.setConfidenceThreshold(FaceDetectConstant.DEFAULT_CONFIDENCE_THRESHOLD);//只返回相似度大于该值的人脸 - config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD);//用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 - config.setDevice(device); + //人脸检测模型，SmartJavaAI提供了多种模型选择(更多模型，请查看文档)，切换模型需要同时修改modelEnum及modelPath + config.setModelEnum(FaceDetModelEnum.MTCNN); + //下载模型并替换本地路径，下载地址：https://pan.baidu.com/s/10l22x5fRz_gwLr8EAHa1Jg?pwd=1234 提取码: 1234 + config.setModelPath("/Users/wenjie/Documents/develop/face_model"); + //只返回相似度大于该值的人脸,需要根据实际情况调整，分值越大越严格容易漏检，分值越小越宽松容易误识别 + config.setConfidenceThreshold(0.5f); + //用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 + config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD); return FaceDetModelFactory.getInstance().getModel(config); } diff --git a/examples/objectdetection-example/pom.xml b/examples/objectdetection-example/pom.xml index bac5870..d5187f1 100644 --- a/examples/objectdetection-example/pom.xml +++ b/examples/objectdetection-example/pom.xml @@ -101,7 +101,7 @@ ai.djl.pytorch pytorch-jni - 2.5.1-0.32.0 + 2.7.1-0.34.0 runtime @@ -138,7 +138,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.windows-x86_64} - 2.5.1 + 2.7.1 runtime @@ -176,7 +176,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.linux-x86_64} - 2.5.1 + 2.7.1 runtime @@ -213,7 +213,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.osx-aarch64} - 2.5.1 + 2.7.1 runtime @@ -251,7 +251,7 @@ ai.djl.pytorch pytorch-native-cpu-precxx11 ${djl.platform.linux-aarch64} - 2.5.1 + 2.7.1 runtime diff --git a/examples/ocr-examples/pom.xml b/examples/ocr-examples/pom.xml index 7d329ff..82dd180 100644 --- a/examples/ocr-examples/pom.xml +++ b/examples/ocr-examples/pom.xml @@ -103,7 +103,7 @@ ai.djl.pytorch pytorch-jni - 2.5.1-0.32.0 + 2.7.1-0.34.0 runtime @@ -140,7 +140,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.windows-x86_64} - 2.5.1 + 2.7.1 runtime @@ -178,7 +178,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.linux-x86_64} - 2.5.1 + 2.7.1 runtime @@ -215,7 +215,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.osx-aarch64} - 2.5.1 + 2.7.1 runtime @@ -253,7 +253,7 @@ ai.djl.pytorch pytorch-native-cpu-precxx11 ${djl.platform.linux-aarch64} - 2.5.1 + 2.7.1 runtime diff --git a/examples/translation-example/pom.xml b/examples/translation-example/pom.xml index d0bafb7..f90843d 100644 --- a/examples/translation-example/pom.xml +++ b/examples/translation-example/pom.xml @@ -101,7 +101,7 @@ ai.djl.pytorch pytorch-jni - 2.5.1-0.32.0 + 2.7.1-0.34.0 runtime @@ -112,7 +112,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.windows-x86_64} - 2.5.1 + 2.7.1 runtime @@ -124,7 +124,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.linux-x86_64} - 2.5.1 + 2.7.1 runtime @@ -134,7 +134,7 @@ ai.djl.pytorch pytorch-native-cpu ${djl.platform.osx-aarch64} - 2.5.1 + 2.7.1 runtime @@ -145,7 +145,7 @@ ai.djl.pytorch pytorch-native-cpu-precxx11 ${djl.platform.linux-aarch64} - 2.5.1 + 2.7.1 runtime diff --git a/face/src/main/java/cn/smartjavaai/face/enums/FaceDetModelEnum.java b/face/src/main/java/cn/smartjavaai/face/enums/FaceDetModelEnum.java index c717736..7994737 100644 --- a/face/src/main/java/cn/smartjavaai/face/enums/FaceDetModelEnum.java +++ b/face/src/main/java/cn/smartjavaai/face/enums/FaceDetModelEnum.java @@ -9,24 +9,12 @@ import lombok.Data; public enum FaceDetModelEnum { RETINA_FACE("PyTorch",0,0, "https://resources.djl.ai/test-models/pytorch/retinaface.zip"), - RETINA_FACE_ONNX("OnnxRuntime",0,0, null), + RETINA_FACE_1080_720_ONNX("OnnxRuntime",1080,720, null), RETINA_FACE_640_ONNX("OnnxRuntime",640,640, null), - RETINA_FACE_320_ONNX("OnnxRuntime",320,320, null), - RETINA_FACE_720_1280_ONNX("OnnxRuntime",720,1280, null), - RETINA_FACE_MOBILE_ONNX("OnnxRuntime",0,0, null), - RETINA_FACE_MOBILE_320_ONNX("OnnxRuntime",320,320, null), - RETINA_FACE_MOBILE_640_ONNX("OnnxRuntime",640,640, null), - RETINA_FACE_MOBILE_720_1280_ONNX("OnnxRuntime",720,1080, null), ULTRA_LIGHT_FAST_GENERIC_FACE("PyTorch",0,0, "https://resources.djl.ai/test-models/pytorch/ultranet.zip"), SEETA_FACE6_MODEL(null,0,0, null), - YOLOV8_FACE("OnnxRuntime",640,640, null), YOLOV5_FACE_640("OnnxRuntime", 640,640, null), YOLOV5_FACE_320("OnnxRuntime",320,320, null), - SCRFD_160("OnnxRuntime",160,160, null), - SCRFD_320("OnnxRuntime",320,320, null), - SCRFD_640("OnnxRuntime",640,640, null), - SCRFD_1280("OnnxRuntime",1280,1280, null), - MTCNN("OnnxRuntime",1280,1280, null); diff --git a/face/src/main/java/cn/smartjavaai/face/factory/FaceDetModelFactory.java b/face/src/main/java/cn/smartjavaai/face/factory/FaceDetModelFactory.java index 4f09aac..1e0f0a2 100644 --- a/face/src/main/java/cn/smartjavaai/face/factory/FaceDetModelFactory.java +++ b/face/src/main/java/cn/smartjavaai/face/factory/FaceDetModelFactory.java @@ -7,6 +7,7 @@ import cn.smartjavaai.face.enums.FaceDetModelEnum; import cn.smartjavaai.face.exception.FaceException; import cn.smartjavaai.face.model.facedect.CommonFaceDetModel; import cn.smartjavaai.face.model.facedect.FaceDetModel; +import cn.smartjavaai.face.model.facedect.MtcnnFaceDetModel; import cn.smartjavaai.face.model.facedect.SeetaFace6FaceDetModel; import cn.smartjavaai.face.model.facerec.*; import lombok.extern.slf4j.Slf4j; @@ -122,11 +123,13 @@ public class FaceDetModelFactory { // 初始化默认算法 static { registerAlgorithm(FaceDetModelEnum.RETINA_FACE, CommonFaceDetModel.class); + registerAlgorithm(FaceDetModelEnum.RETINA_FACE_640_ONNX, CommonFaceDetModel.class); + registerAlgorithm(FaceDetModelEnum.RETINA_FACE_1080_720_ONNX, CommonFaceDetModel.class); registerAlgorithm(FaceDetModelEnum.ULTRA_LIGHT_FAST_GENERIC_FACE, CommonFaceDetModel.class); registerAlgorithm(FaceDetModelEnum.SEETA_FACE6_MODEL, SeetaFace6FaceDetModel.class); - registerAlgorithm(FaceDetModelEnum.YOLOV8_FACE, CommonFaceDetModel.class); registerAlgorithm(FaceDetModelEnum.YOLOV5_FACE_640, CommonFaceDetModel.class); registerAlgorithm(FaceDetModelEnum.YOLOV5_FACE_320, CommonFaceDetModel.class); + registerAlgorithm(FaceDetModelEnum.MTCNN, MtcnnFaceDetModel.class); log.debug("缓存目录：{}", Config.getCachePath()); } diff --git a/face/src/main/java/cn/smartjavaai/face/model/facedect/MtcnnFaceDetModel.java b/face/src/main/java/cn/smartjavaai/face/model/facedect/MtcnnFaceDetModel.java index 673594f..376f999 100644 --- a/face/src/main/java/cn/smartjavaai/face/model/facedect/MtcnnFaceDetModel.java +++ b/face/src/main/java/cn/smartjavaai/face/model/facedect/MtcnnFaceDetModel.java @@ -19,10 +19,7 @@ import cn.smartjavaai.common.entity.*; import cn.smartjavaai.common.entity.Point; import cn.smartjavaai.common.entity.face.FaceInfo; import cn.smartjavaai.common.pool.PredictorFactory; -import cn.smartjavaai.common.utils.Base64ImageUtils; -import cn.smartjavaai.common.utils.FileUtils; -import cn.smartjavaai.common.utils.ImageUtils; -import cn.smartjavaai.common.utils.OpenCVUtils; +import cn.smartjavaai.common.utils.*; import cn.smartjavaai.face.config.FaceDetConfig; import cn.smartjavaai.face.exception.FaceException; import cn.smartjavaai.face.model.facedect.criterial.FaceDetCriteriaFactory; @@ -85,8 +82,8 @@ public class MtcnnFaceDetModel implements FaceDetModel{ Path rnetPath = modelPath.resolve("rnet_script.pt"); Path onetPath = modelPath.resolve("onet_script.pt"); pNetModel = getModel(pnetPath); - rNetModel = getModel(pnetPath); - oNetModel = getModel(pnetPath); + rNetModel = getModel(rnetPath); + oNetModel = getModel(onetPath); this.pnetPredictorPool = new GenericObjectPool<>(new PredictorFactory<>(pNetModel)); this.rnetPredictorPool = new GenericObjectPool<>(new PredictorFactory<>(rNetModel)); @@ -266,22 +263,84 @@ public class MtcnnFaceDetModel implements FaceDetModel{ * @return */ public R detect(Image image){ - try (NDManager manager = NDManager.newBaseManager(pNetModel.getNDManager().getDevice())){ - List scales = MtcnnProcess.generateScales(image); - NDArray imgs = MtcnnProcess.processInput(manager, image); + Predictor pNetPredictor = null; + Predictor rNetPredictor = null; + Predictor oNetPredictor = null; + try (NDManager manager = pNetModel.getNDManager().newSubManager();){ + pNetPredictor = pnetPredictorPool.borrowObject(); + rNetPredictor = rnetPredictorPool.borrowObject(); + oNetPredictor = onetPredictorPool.borrowObject(); int h = image.getHeight(); int w = image.getWidth(); - NDList outputPnet = PNetModel.firstStage(manager, pnetPredictorPool.borrowObject(), imgs, scales, w, h); + //第一阶段 + NDList outputPnet = PNetModel.firstStage(manager, pNetPredictor, image); + if(CollectionUtils.isEmpty(outputPnet)){ + return R.fail(R.Status.NO_FACE_DETECTED); + } NDArray boxes = outputPnet.get(0); NDArray image_inds = outputPnet.get(1); + NDArray imgs = outputPnet.get(2); + if(DJLCommonUtils.isNDArrayEmpty(boxes) || DJLCommonUtils.isNDArrayEmpty(image_inds) || DJLCommonUtils.isNDArrayEmpty(imgs)){ + return R.fail(R.Status.NO_FACE_DETECTED); + } NDList pad = MtcnnUtils.pad(boxes, w, h); - NDList outputRnet = RNetModel.secondStage(manager, rnetPredictorPool.borrowObject(), imgs,boxes,pad, image_inds); + //第二阶段 + NDList outputRnet = RNetModel.secondStage(manager, rNetPredictor, imgs,boxes,pad, image_inds); + if(CollectionUtils.isEmpty(outputRnet)){ + return R.fail(R.Status.NO_FACE_DETECTED); + } NDArray image_indsFiltered = outputRnet.get(0); NDArray scoresFiltered = outputRnet.get(1); - MtcnnBatchResult oNetResult = ONetModel.thirdStage(manager, onetPredictorPool.borrowObject(), imgs,boxes, w, h, scoresFiltered, image_indsFiltered); - return R.ok(convertToDetectionResponse(oNetResult)); + boxes = outputRnet.get(2); + if(DJLCommonUtils.isNDArrayEmpty(boxes) || DJLCommonUtils.isNDArrayEmpty(image_indsFiltered) || DJLCommonUtils.isNDArrayEmpty(scoresFiltered)){ + return R.fail(R.Status.NO_FACE_DETECTED); + } + //第三阶段 + MtcnnBatchResult oNetResult = ONetModel.thirdStage(manager, oNetPredictor, imgs,boxes, w, h, scoresFiltered, image_indsFiltered); + DetectionResponse detectionResponse = convertToDetectionResponse(oNetResult); + if(Objects.isNull(detectionResponse)){ + return R.fail(R.Status.NO_FACE_DETECTED); + } + return R.ok(detectionResponse); } catch (Exception e) { throw new RuntimeException(e); + } finally { + if (pNetPredictor != null) { + try { + pnetPredictorPool.returnObject(pNetPredictor); //归还 + } catch (Exception e) { + log.warn("归还Predictor失败", e); + try { + pNetPredictor.close(); // 归还失败才销毁 + } catch (Exception ex) { + log.error("关闭Predictor失败", ex); + } + } + } + if (rNetPredictor != null) { + try { + rnetPredictorPool.returnObject(rNetPredictor); //归还 + } catch (Exception e) { + log.warn("归还Predictor失败", e); + try { + rNetPredictor.close(); // 归还失败才销毁 + } catch (Exception ex) { + log.error("关闭Predictor失败", ex); + } + } + } + if (oNetPredictor != null) { + try { + onetPredictorPool.returnObject(oNetPredictor); //归还 + } catch (Exception e) { + log.warn("归还Predictor失败", e); + try { + oNetPredictor.close(); // 归还失败才销毁 + } catch (Exception ex) { + log.error("关闭Predictor失败", ex); + } + } + } } } @@ -306,6 +365,9 @@ public class MtcnnFaceDetModel implements FaceDetModel{ NDArray probs = mtcnnBatchResult.probs.get(0); NDArray points = mtcnnBatchResult.points.get(0); + if (DJLCommonUtils.isNDArrayEmpty(boxes) || DJLCommonUtils.isNDArrayEmpty(probs) || DJLCommonUtils.isNDArrayEmpty(points)){ + return null; + } long numBoxes = boxes.getShape().get(0); for (int i = 0; i < numBoxes; i++) { float[] boxCoords = boxes.get(i).toFloatArray(); // [x1, y1, x2, y2] diff --git a/face/src/main/java/cn/smartjavaai/face/model/facedect/criterial/FaceDetCriteriaFactory.java b/face/src/main/java/cn/smartjavaai/face/model/facedect/criterial/FaceDetCriteriaFactory.java index a8b0bd9..4a8bb9b 100644 --- a/face/src/main/java/cn/smartjavaai/face/model/facedect/criterial/FaceDetCriteriaFactory.java +++ b/face/src/main/java/cn/smartjavaai/face/model/facedect/criterial/FaceDetCriteriaFactory.java @@ -8,6 +8,7 @@ import ai.djl.repository.zoo.Criteria; import ai.djl.training.util.ProgressBar; import ai.djl.translate.Translator; import cn.smartjavaai.common.enums.DeviceEnum; +import cn.smartjavaai.common.utils.LetterBoxUtils; import cn.smartjavaai.face.config.FaceDetConfig; import cn.smartjavaai.face.config.FaceExpressionConfig; import cn.smartjavaai.face.constant.FaceDetectConstant; @@ -19,9 +20,7 @@ import cn.smartjavaai.face.exception.FaceException; import cn.smartjavaai.face.model.expression.translator.DenseNetEmotionTranslator; import cn.smartjavaai.face.model.expression.translator.FrEmotionTranslator; import cn.smartjavaai.face.translator.FaceDetectionTranslator; -import cn.smartjavaai.face.translator.SCRFDFaceTranslator; import cn.smartjavaai.face.translator.YoloV5FaceTranslator; -import cn.smartjavaai.face.translator.YoloV8FaceTranslator; import org.apache.commons.lang3.StringUtils; import java.nio.file.Paths; @@ -73,33 +72,19 @@ public class FaceDetCriteriaFactory { */ public static Translator getTranslator(FaceDetConfig config) { Translator translator = null; - if(config.getModelEnum() == FaceDetModelEnum.RETINA_FACE){ + if(config.getModelEnum() == FaceDetModelEnum.RETINA_FACE || config.getModelEnum() == FaceDetModelEnum.RETINA_FACE_640_ONNX + || config.getModelEnum() == FaceDetModelEnum.RETINA_FACE_1080_720_ONNX){ translator = - new FaceDetectionTranslator(config.getConfidenceThreshold(), config.getNmsThresh(), RetinaFaceConstant.variance, FaceDetectConstant.MAX_FACE_LIMIT, RetinaFaceConstant.scales, RetinaFaceConstant.steps); + new FaceDetectionTranslator(config.getConfidenceThreshold(), config.getNmsThresh(), RetinaFaceConstant.variance, FaceDetectConstant.MAX_FACE_LIMIT, + RetinaFaceConstant.scales, RetinaFaceConstant.steps, config.getModelEnum().getInputWidth(), config.getModelEnum().getInputHeight()); }else if (config.getModelEnum() == FaceDetModelEnum.ULTRA_LIGHT_FAST_GENERIC_FACE){ translator = - new FaceDetectionTranslator(config.getConfidenceThreshold(), config.getNmsThresh(), UltraLightFastGenericFaceConstant.variance, FaceDetectConstant.MAX_FACE_LIMIT, UltraLightFastGenericFaceConstant.scales, UltraLightFastGenericFaceConstant.steps); - }else if (config.getModelEnum() == FaceDetModelEnum.YOLOV8_FACE){ - Map arguments = new HashMap<>(); - arguments.put("width", 640); - arguments.put("height", 640); - arguments.put("resizeShort", 640); - arguments.put("centerFit", true); - translator = YoloV8FaceTranslator.builder(arguments).build(); + new FaceDetectionTranslator(config.getConfidenceThreshold(), config.getNmsThresh(), UltraLightFastGenericFaceConstant.variance, + FaceDetectConstant.MAX_FACE_LIMIT, UltraLightFastGenericFaceConstant.scales, UltraLightFastGenericFaceConstant.steps, config.getModelEnum().getInputWidth(), config.getModelEnum().getInputHeight()); }else if (config.getModelEnum() == FaceDetModelEnum.YOLOV5_FACE_640 || config.getModelEnum() == FaceDetModelEnum.YOLOV5_FACE_320){ - Map arguments = new HashMap<>(); - arguments.put("width", config.getModelEnum().getInputWidth()); - arguments.put("height", config.getModelEnum().getInputHeight()); - arguments.put("resizeShort", true); - arguments.put("centerFit", true); - translator = YoloV5FaceTranslator.builder(arguments).build(); - }else if (config.getModelEnum() == FaceDetModelEnum.SCRFD_160 - || config.getModelEnum() == FaceDetModelEnum.SCRFD_320 - || config.getModelEnum() == FaceDetModelEnum.SCRFD_640 - || config.getModelEnum() == FaceDetModelEnum.SCRFD_1280){ - translator = - new SCRFDFaceTranslator(config.getConfidenceThreshold(), config.getNmsThresh(), 100, new int[]{8, 16, 32}); + translator = YoloV5FaceTranslator.builder() + .setImageSize(config.getModelEnum().getInputWidth(), config.getModelEnum().getInputWidth()).build(); } return translator; } diff --git a/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/ONetModel.java b/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/ONetModel.java index fcf94fd..4fa2b43 100644 --- a/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/ONetModel.java +++ b/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/ONetModel.java @@ -65,7 +65,7 @@ public class ONetModel { NDArray out0 = out.get(0).transpose(1, 0); // permute(1,0) NDArray out1 = out.get(1).transpose(1, 0); NDArray out2 = out.get(2).transpose(1, 0); - NDArray score = out1.get(1); // out1[1, :] + NDArray score = out2.get(1); // out1[1, :] points = out1.duplicate(); NDArray ipass = score.gt(0.7); // score > threshold[1] // ipass 为布尔/0-1张量，长度应等于 points 的第 1 维（这里是 7） @@ -84,7 +84,7 @@ public class ONetModel { NDArray boxesSelected = boxes.get(manager.create(validIndices)); // 行筛选 // 取前 4 列 boxesSelected = boxesSelected.get(":, 0:4"); // 只保留前 4 列 - scoresFiltered = scoresFiltered.get(ipass).reshape(-1, 1); // score[ipass].unsqueeze(1) + scoresFiltered = score.get(ipass).reshape(-1, 1); // score[ipass].unsqueeze(1) boxes = NDArrays.concat(new NDList(boxesSelected, scoresFiltered), 1); // 拼接成 (N,5) // 筛选 image_inds @@ -92,7 +92,6 @@ public class ONetModel { NDArray mv = out0.transpose() // (N, 4) .get(ipass); // 1-D 花式索引在第 0 维，得到 (k, 4) - System.out.println("----------"); // w_i = boxes[:, 2] - boxes[:, 0] + 1 NDArray w_i = boxes.get(":,2").sub(boxes.get(":,0")).add(1); diff --git a/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/PNetModel.java b/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/PNetModel.java index 456e9ad..2dd3a82 100644 --- a/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/PNetModel.java +++ b/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/PNetModel.java @@ -23,53 +23,80 @@ import java.util.List; public class PNetModel { - /** - * 输入图片 - * @param imgs - * @param w - * @param h - * @param scale + * 生成金字塔缩放比例列表 + * @param image * @return */ - public static NDList processInput(NDArray imgs, int w, int h, double scale){ - int newH = (int) (h * scale + 1); - int newW = (int) (w * scale + 1); - // (N, H, W, C) - NDArray transposed = imgs.transpose(0, 2, 3, 1); - transposed = NDImageUtils.resize(transposed, newW, newH, Image.Interpolation.AREA); - // (N, C, H, W) - transposed = transposed.transpose(0, 3, 1, 2); - // 归一化 - transposed = transposed.sub(127.5).mul(0.0078125f); - System.out.println("inputPnet: " + transposed.getShape()); - return new NDList(transposed); - } + public static List generateScales(Image image){ + long h = image.getHeight(); + long w = image.getWidth(); + // 计算最小缩放比例 + double minsize = 20; + double m = 12.0 / minsize; + double minl = Math.min(h, w) * m; - public static NDList processOutput(NDList outputPnet, double scale, NDManager manager){ - NDArray reg = outputPnet.get(0); // [B, 4, H, W] - NDArray probs = outputPnet.get(1); // [B, 2, H, W] - List boundingBox = generateBoundingBox(reg, probs.get(":, 1"), (float)scale, 0.6f); - NDArray boxes_scale = boundingBox.get(0); // [N,9] - NDArray imgIndND = boundingBox.get(1); // [N] - NDArray pick = NMSUtils.batchedNms(boxes_scale.get(":,:4"), boxes_scale.get(":,4"), imgIndND, 0.5f, manager); - return new NDList(boxes_scale, imgIndND, pick); + // 创建金字塔缩放比例列表 + double factor = 0.709; // 你原代码的 factor + List scales = new ArrayList<>(); + double scale_i = m; + while (minl >= 12) { + scales.add(scale_i); + scale_i *= factor; + minl *= factor; + } + return scales; } - public static NDList firstStage(NDManager manager, Predictor pnetPredictor, NDArray imgs, List scales, int width, int height) throws TranslateException { + public static NDArray pNetPre(Image input,NDManager manager){ + // Image -> NDArray (H, W, C) + NDArray array = input.toNDArray(manager, Image.Flag.COLOR); + // 增加 batch 维度 (1, H, W, C) ---- + array = array.expandDims(0); + // 交换维度 (N, C, H, W) + array = array.transpose(0, 3, 1, 2); + // 转成模型的数据类型 + if (!array.getDataType().equals(DataType.FLOAT32)) { + array = array.toType(DataType.FLOAT32, false); + } + return array; + } + + + public static NDList firstStage(NDManager manager, Predictor pnetPredictor, Image image) throws TranslateException { + List scales = MtcnnProcess.generateScales(image); + NDArray imgs = pNetPre(image,manager); + int h = image.getHeight(); + int w = image.getWidth(); // 第一阶段 NDList boxes_list = new NDList(); NDList image_inds_list = new NDList(); NDList scale_picks_list = new NDList(); int offset = 0; for (double scale : scales) { - NDList inputPnet = processInput(imgs, width, height, scale); - NDList outputPnet = pnetPredictor.predict(inputPnet); - NDList output = processOutput(outputPnet, scale, manager); - NDArray boxes_scale = output.get(0); - NDArray imgIndND = output.get(1); - NDArray pick = output.get(2); + int newH = (int) (h * scale + 1); + int newW = (int) (w * scale + 1); + // (N, H, W, C) + NDArray transposed = imgs.transpose(0, 2, 3, 1); + transposed = NDImageUtils.resize(transposed, newW, newH, Image.Interpolation.AREA); + // (N, C, H, W) + transposed = transposed.transpose(0, 3, 1, 2); + // 归一化 + transposed = transposed.sub(127.5).mul(0.0078125f); + + + NDList outputPnet = pnetPredictor.predict(new NDList(transposed)); + NDArray reg = outputPnet.get(0); // [B, 4, H, W] + NDArray probs = outputPnet.get(1); // [B, 2, H, W] + + List boundingBox = generateBoundingBox(reg, probs.get(":, 1"), (float)scale, 0.6f); + + NDArray boxes_scale = boundingBox.get(0); // [N,9] + NDArray imgIndND = boundingBox.get(1); // [N] + + NDArray pick = NMSUtils.batchedNms(boxes_scale.get(":,:4"), boxes_scale.get(":,4"), imgIndND, 0.5f,manager); + boxes_list.add(boxes_scale); image_inds_list.add(imgIndND); scale_picks_list.add(pick.add(offset)); @@ -96,7 +123,6 @@ public class PNetModel { boxes = boxes.get(pick); image_inds = image_inds.get(pick); - System.out.println(Arrays.toString(boxes.get(0).toFloatArray())); NDArray regw = boxes.get(":, 2").sub(boxes.get(":, 0")); NDArray regh = boxes.get(":, 3").sub(boxes.get(":, 1")); @@ -108,7 +134,7 @@ public class PNetModel { boxes = NDArrays.stack(new NDList(qq1, qq2, qq3, qq4, boxes.get(":, 4")), 1); boxes = MtcnnUtils.rerec(boxes); - return new NDList(boxes, image_inds); + return new NDList(boxes, image_inds, imgs); } /** @@ -131,13 +157,8 @@ public class PNetModel { // mask = probs >= thresh -> [B,H,W] NDArray mask = probs.gte(threshold); - System.out.println("reg: " + reg.getShape()); - System.out.println("probs shape: " + probs.getShape()); - System.out.println("scale: " + scale); - System.out.println("mask: " + mask.getShape()); // mask_inds = mask.nonzero() -> [N,3] 每行: (batch, y, x) NDArray maskInds = mask.nonzero(); - System.out.println("maskInds: " + maskInds.getShape()); // image_inds = mask_inds[:, 0] NDArray imageInds = maskInds.get(":,0"); diff --git a/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/RNetModel.java b/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/RNetModel.java index 2f4aa2a..9d4e1a5 100644 --- a/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/RNetModel.java +++ b/face/src/main/java/cn/smartjavaai/face/model/facedect/mtcnn/RNetModel.java @@ -101,7 +101,7 @@ public class RNetModel { log.debug("No face detected."); return null; } - return new NDList(image_indsFiltered, scoresFiltered); + return new NDList(image_indsFiltered, scoresFiltered,boxes); } } diff --git a/face/src/main/java/cn/smartjavaai/face/translator/FaceDetectionTranslator.java b/face/src/main/java/cn/smartjavaai/face/translator/FaceDetectionTranslator.java index 341524e..4c71596 100644 --- a/face/src/main/java/cn/smartjavaai/face/translator/FaceDetectionTranslator.java +++ b/face/src/main/java/cn/smartjavaai/face/translator/FaceDetectionTranslator.java @@ -13,6 +13,7 @@ package cn.smartjavaai.face.translator; import ai.djl.modality.cv.Image; +import ai.djl.modality.cv.ImageFactory; import ai.djl.modality.cv.output.*; import ai.djl.ndarray.NDArray; import ai.djl.ndarray.NDArrays; @@ -22,7 +23,11 @@ import ai.djl.ndarray.types.DataType; import ai.djl.ndarray.types.Shape; import ai.djl.translate.Translator; import ai.djl.translate.TranslatorContext; +import cn.smartjavaai.common.utils.LetterBoxUtils; +import java.io.IOException; +import java.nio.file.Files; +import java.nio.file.Paths; import java.util.ArrayList; import java.util.List; import java.util.Map; @@ -40,29 +45,50 @@ public class FaceDetectionTranslator implements Translator 0 && inputHeight > 0){ + //Letter box resize 640x640 with padding (保持比例，补边缘) + LetterBoxUtils.ResizeResult letterBoxResult = LetterBoxUtils.letterbox(ctx.getNDManager(), array, inputWidth, inputHeight, 114f, LetterBoxUtils.PaddingPosition.CENTER); + array = letterBoxResult.image; + ctx.setAttachment("needRecover", "1");//需要还原 + ctx.setAttachment("scale", letterBoxResult.r); + }else{ + ctx.setAttachment("needRecover", "0");//不需要还原 + ctx.setAttachment("width", input.getWidth()); + ctx.setAttachment("height", input.getHeight()); + } + array = array.transpose(2, 0, 1).flip(0); // HWC -> CHW RGB -> BGR // The network by default takes float32 if (!array.getDataType().equals(DataType.FLOAT32)) { @@ -80,6 +106,13 @@ public class FaceDetectionTranslator implements Translator keyPoints = new ArrayList<>(); for (int j = 0; j < 5; j++) { // 5 face landmarks double x = landmsArr[j * 2]; @@ -150,9 +183,14 @@ public class FaceDetectionTranslator implements Translator { - - private double confThresh; - private double nmsThresh; - private int topK; - private int[] steps; - private int inputWidth = 640; - private int inputHeight = 640; - - public SCRFDFaceTranslator( - double confThresh, - double nmsThresh, - int topK, - int[] steps) { - this.confThresh = confThresh; - this.nmsThresh = nmsThresh; - this.topK = topK; - this.steps = steps; - } - - /** {@inheritDoc} */ - @Override - public NDList processInput(TranslatorContext ctx, Image input) { - - ctx.setAttachment("width", input.getWidth()); - ctx.setAttachment("height", input.getHeight()); - - NDManager manager = ctx.getNDManager(); - NDArray array = input.toNDArray(manager, Image.Flag.COLOR); - - //Letter box resize 640x640 with padding (保持比例，补边缘) - LetterBoxUtils.ResizeResult letterBoxResult = LetterBoxUtils.letterbox(manager, array, inputWidth, inputHeight, 0f, LetterBoxUtils.PaddingPosition.LEFT_TOP); - ctx.setAttachment("scale", letterBoxResult.r); - array = letterBoxResult.image; - array = array.transpose(2, 0, 1).flip(0); // HWC -> CHW RGB -> BGR - // The network by default takes float32 - if (!array.getDataType().equals(DataType.FLOAT32)) { - array = array.toType(DataType.FLOAT32, false); - } - // 归一化 - array = array.sub(127.5).mul(0.0078125); - return new NDList(array); - } - - /** {@inheritDoc} */ - @Override - public DetectedObjects processOutput(TranslatorContext ctx, NDList list) { - - NDManager manager = - NDManager.newBaseManager(ctx.getNDManager().getDevice(), "PyTorch"); - int sourceWidth = (int) ctx.getAttachment("width"); - int sourceHeight = (int) ctx.getAttachment("height"); - float detScale = (float) ctx.getAttachment("scale"); - - Map centerCache = new HashMap<>(); - List scores_list = new ArrayList<>(); - List bboxes_list = new ArrayList<>(); - List kpss_list = new ArrayList<>(); - // 步长长度 - int fmc = steps.length; - int numAnchors = 2; - // 多尺度后处理 - for (int idx = 0; idx < steps.length; idx++) { - int stride = steps[idx]; - NDArray scores, bboxPreds, kpsPreds = null; - scores = list.get(idx); - bboxPreds = list.get(idx + fmc).mul(stride); - kpsPreds = list.get(idx + fmc * 2).mul(stride); - int height = inputHeight / stride; - int width = inputWidth / stride; - String key = height + "_" + width + "_" + stride; - // anchor centers cache - NDArray anchorCenters; - if (centerCache.containsKey(key)) { - anchorCenters = centerCache.get(key); - } else { - NDArray yv = manager.arange((float) height).reshape(height, 1).repeat(1, width); - NDArray xv = manager.arange((float) width).reshape(1, width).repeat(0, height); - // stack x, y 到最后一维 - anchorCenters = xv.stack(yv, -1); // shape [height, width, 2] - // 乘 stride - anchorCenters = anchorCenters.mul(stride); - // 拉平成 [-1, 2]，等价于 NumPy 的 reshape((-1, 2)) - long total = anchorCenters.getShape().get(0) * anchorCenters.getShape().get(1); - // anchorCenters 现在是 [height*width, 2] - anchorCenters = anchorCenters.reshape(total, 2); - // 在第一维重复 numAnchors 次，直接拉平成最终形状 - anchorCenters = anchorCenters.repeat(0, numAnchors); // shape [N*numAnchors, 2] - if (centerCache.size() < 100) { - centerCache.put(key, anchorCenters); - } - } - - NDArray pos_mask = scores.gte(confThresh); // scores >= thresh - NDArray pos_inds = pos_mask.nonzero(); // shape: [N, 2] - pos_inds = pos_inds.get(":, 0"); // 取第一列的索引 -// System.out.println(Arrays.toString(pos_inds.toLongArray())); - // 计算 bbox - NDArray bboxes = distance2bbox(anchorCenters, bboxPreds); // [num_anchors, 4] - - // 取出符合阈值的 - NDArray pos_scores = scores.get(pos_inds); - NDArray pos_bboxes = bboxes.get(pos_inds); - - scores_list.add(pos_scores); - bboxes_list.add(pos_bboxes); - - NDArray kpss = distance2kps(anchorCenters, kpsPreds); // [num_anchors, num_kps*2] - kpss = kpss.reshape(kpss.getShape().get(0), -1, 2); // reshape (N, -1, 2) - NDArray pos_kpss = kpss.get(pos_inds); - kpss_list.add(pos_kpss); - } - - // 1. 合并 scores - NDArray scores = NDArrays.concat(new NDList(scores_list), 0); - NDArray scoresRavel = scores.reshape(-1); - - // 2. 得到排序索引 - long[] orderLong = scoresRavel.argSort().flip(0).get(":" + topK).toLongArray(); - - // 3. 合并 bboxes - NDArray bboxes = NDArrays.concat(new NDList(bboxes_list), 0).div(detScale); - - NDArray kpss = NDArrays.concat(new NDList(kpss_list), 0).div(detScale); - // 4. 拼接 [x1,y1,x2,y2,score] - NDArray preDet = bboxes.concat(scores.reshape(-1,1), 1); - // 5. 按 order 排序 - preDet = preDet.get(manager.create(orderLong)); - // 6. NMS - int[] keep = NMSUtils.nms(preDet.get(":,0:4"), preDet.get(":,4"), (float)nmsThresh); - NDArray det = preDet.get(manager.create(keep)); -// System.out.println(Arrays.toString(det.toFloatArray())); - if (kpss != null) { - kpss = kpss.get(manager.create(orderLong)); - kpss = kpss.get(manager.create(keep)); - } - List retNames = new ArrayList<>(); - List retProbs = new ArrayList<>(); - List retBB = new ArrayList<>(); - long numDet = det.getShape().get(0); // N - long numCols = det.getShape().get(1); // 应该是 5: x1,y1,x2,y2,score - float[] flat = det.toFloatArray(); // 一维 - for (int i = 0; i < numDet; i++) { - int base = (int) (i * numCols); - float x1 = flat[base] / sourceWidth; - float y1 = flat[base + 1] / sourceHeight; - float x2 = flat[base + 2] / sourceWidth; - float y2 = flat[base + 3] / sourceHeight; - float score = flat[base + 4]; - - retNames.add("face"); // 类别 - retProbs.add((double) score); - - float width = x2 - x1; - float height = y2 - y1; - Landmark rect = new Landmark(x1, y1, width, height, decodeKps(kpss.get(i))); - retBB.add(rect); - } - return new DetectedObjects(retNames, retProbs, retBB); - } - - - public NDArray distance2bbox(NDArray points, NDArray distance) { - // points: [N, 2], distance: [N, 4] - NDArray x1 = points.get(":, 0").sub(distance.get(":, 0")); // x - left - NDArray y1 = points.get(":, 1").sub(distance.get(":, 1")); // y - top - NDArray x2 = points.get(":, 0").add(distance.get(":, 2")); // x + right - NDArray y2 = points.get(":, 1").add(distance.get(":, 3")); // y + bottom - // stack([x1, y1, x2, y2], axis=-1) - NDList list = new NDList(x1.expandDims(1), y1.expandDims(1), x2.expandDims(1), y2.expandDims(1)); - return NDArrays.concat(list, 1); // axis=1 表示最后一维 - } - - - public NDArray distance2kps(NDArray points, NDArray distance) { - // points: [N, 2], distance: [N, 2*num_kps] - int numKps = (int) distance.getShape().get(1) / 2; - List preds = new ArrayList<>(); - for (int i = 0; i < numKps * 2; i += 2) { - NDArray px = points.get(":, " + (i % 2)).add(distance.get(":, " + i)); - NDArray py = points.get(":, " + ((i % 2) + 1)).add(distance.get(":, " + (i + 1))); - preds.add(px); - preds.add(py); - } - // stack(preds, axis=-1) - NDList stackList = new NDList(); - for (NDArray arr : preds) { - stackList.add(arr.expandDims(1)); - } - return NDArrays.concat(stackList, 1); // shape [N, num_kps*2] - } - - - public List decodeKps(NDArray kpss) { - // 转成一维 float 数组 - float[] flat = kpss.toFloatArray(); - - // reshape 成二维 [5][2] - int numPoints = (int) kpss.getShape().get(0); // 5 - int dim = (int) kpss.getShape().get(1); // 2 - float[][] kpsArray = new float[numPoints][dim]; - for (int i = 0; i < numPoints; i++) { - for (int j = 0; j < dim; j++) { - kpsArray[i][j] = flat[i * dim + j]; - } - } - - // 转成 Point 数组 - List points = new ArrayList<>(); - for (int i = 0; i < numPoints; i++) { - points.add(new Point(Math.round(kpsArray[i][0]), Math.round(kpsArray[i][1]))); - } - return points; - } - - -} diff --git a/face/src/main/java/cn/smartjavaai/face/translator/YoloV5FaceTranslator.java b/face/src/main/java/cn/smartjavaai/face/translator/YoloV5FaceTranslator.java index 8733de6..0f41abe 100644 --- a/face/src/main/java/cn/smartjavaai/face/translator/YoloV5FaceTranslator.java +++ b/face/src/main/java/cn/smartjavaai/face/translator/YoloV5FaceTranslator.java @@ -8,6 +8,7 @@ import ai.djl.ndarray.NDList; import ai.djl.ndarray.types.DataType; import ai.djl.ndarray.types.Shape; import ai.djl.translate.*; +import cn.smartjavaai.common.utils.LetterBoxUtils; import java.util.*; @@ -79,8 +80,59 @@ public class YoloV5FaceTranslator implements Translator return builder; } + @Override + public NDList processInput(TranslatorContext ctx, Image input) throws Exception { + NDArray array = input.toNDArray(ctx.getNDManager(), flag); +// NDList list = pipeline.transform(new NDList(array)); +// Shape shape = list.get(0).getShape(); +// int processedWidth; +// int processedHeight; +// long[] dim = shape.getShape(); +// if (NDImageUtils.isCHW(shape)) { +// processedWidth = (int) dim[dim.length - 1]; +// processedHeight = (int) dim[dim.length - 2]; +// } else { +// processedWidth = (int) dim[dim.length - 2]; +// processedHeight = (int) dim[dim.length - 3]; +// } + + LetterBoxUtils.ResizeResult letterBoxResult = LetterBoxUtils.letterbox(ctx.getNDManager(), array, width, height, 114f, LetterBoxUtils.PaddingPosition.CENTER); + array = letterBoxResult.image; + ctx.setAttachment("width", input.getWidth()); + ctx.setAttachment("height", input.getHeight()); + ctx.setAttachment("processedWidth", width); + ctx.setAttachment("processedHeight", height); + ctx.setAttachment("scale", letterBoxResult.r); + // 转为 float32 且归一化到 0~1 + array = array.toType(DataType.FLOAT32, false).div(255f); // HWC + // HWC -> CHW + array = array.transpose(2, 0, 1); // CHW +// return new NDList(array.expandDims(0)); + return new NDList(array); + } + + @Override + public DetectedObjects processOutput(TranslatorContext ctx, NDList list) throws Exception { + int imageWidth = (Integer) ctx.getAttachment("width"); + int imageHeight = (Integer) ctx.getAttachment("height"); + float scale = (Float) ctx.getAttachment("scale"); + switch (yoloOutputLayerType) { + case DETECT: + return processFromDetectOutput(); + case AUTO: + if (list.get(0).getShape().dimension() > 2) { + return processFromDetectOutput(); + } else { + return processFromBoxOutput(imageWidth, imageHeight, list, scale); + } + case BOX: + default: + return processFromBoxOutput(imageWidth, imageHeight, list, scale); + } + } + /** {@inheritDoc} */ - protected DetectedObjects processFromBoxOutput(int imageWidth, int imageHeight, NDList list) { + protected DetectedObjects processFromBoxOutput(int imageWidth, int imageHeight, NDList list, float scale) { float[] flattened = list.get(0).toFloatArray(); int sizeClasses = classes.size(); int stride = 15 + sizeClasses; @@ -119,7 +171,7 @@ public class YoloV5FaceTranslator implements Translator classIds.add(maxIndex); } } - return nms(imageWidth, imageHeight, boxes, classIds, scores); + return nms(imageWidth, imageHeight, boxes, classIds, scores, scale); } private DetectedObjects processFromDetectOutput() { @@ -127,53 +179,16 @@ public class YoloV5FaceTranslator implements Translator "detect layer output is not supported yet, check correct YoloV5 export format"); } - @Override - public DetectedObjects processOutput(TranslatorContext ctx, NDList list) throws Exception { - int imageWidth = (Integer) ctx.getAttachment("width"); - int imageHeight = (Integer) ctx.getAttachment("height"); - switch (yoloOutputLayerType) { - case DETECT: - return processFromDetectOutput(); - case AUTO: - if (list.get(0).getShape().dimension() > 2) { - return processFromDetectOutput(); - } else { - return processFromBoxOutput(imageWidth, imageHeight, list); - } - case BOX: - default: - return processFromBoxOutput(imageWidth, imageHeight, list); - } - } - @Override - public NDList processInput(TranslatorContext ctx, Image input) throws Exception { - NDArray array = input.toNDArray(ctx.getNDManager(), flag); - NDList list = pipeline.transform(new NDList(array)); - Shape shape = list.get(0).getShape(); - int processedWidth; - int processedHeight; - long[] dim = shape.getShape(); - if (NDImageUtils.isCHW(shape)) { - processedWidth = (int) dim[dim.length - 1]; - processedHeight = (int) dim[dim.length - 2]; - } else { - processedWidth = (int) dim[dim.length - 2]; - processedHeight = (int) dim[dim.length - 3]; - } - ctx.setAttachment("width", input.getWidth()); - ctx.setAttachment("height", input.getHeight()); - ctx.setAttachment("processedWidth", processedWidth); - ctx.setAttachment("processedHeight", processedHeight); - return list; - } + + protected DetectedObjects nms( int imageWidth, int imageHeight, List boxes, List classIds, - List scores) { + List scores, float scale) { List retClasses = new ArrayList<>(); List retProbs = new ArrayList<>(); List retBB = new ArrayList<>(); @@ -196,34 +211,41 @@ public class YoloV5FaceTranslator implements Translator for (int index : nms) { int pos = map.get(index); int id = classIds.get(pos); - retClasses.add(classes.get(id)); + + int percent = (int) Math.round(scores.get(pos).doubleValue() * 100); + String className = "face " + percent + "%"; // classes.get(classId) + retClasses.add(className); retProbs.add(scores.get(pos).doubleValue()); // Rectangle rect = boxes.get(pos); Landmark rect = boxes.get(pos); List keypoints = new ArrayList<>(); - if (removePadding) { - int padW = (width - imageWidth) / 2; - int padH = (height - imageHeight) / 2; - rect.getPath().forEach(point -> { - keypoints.add(new Point(point.getX() - padW, point.getY() - padH)); - }); - rect = - new Landmark( - (rect.getX() - padW) / imageWidth, - (rect.getY() - padH) / imageHeight, - rect.getWidth() / imageWidth, - rect.getHeight() / imageHeight,keypoints); - } else if (applyRatio) { - rect.getPath().forEach(point -> { - keypoints.add(new Point(point.getX() / width, point.getY() / height)); - }); - rect = - new Landmark( - rect.getX() / width, - rect.getY() / height, - rect.getWidth() / width, - rect.getHeight() / height,keypoints); - } + + //恢复原图坐标（除回比例，减掉 padding） + rect = LetterBoxUtils.restoreBox(rect, scale, imageWidth, imageHeight, width, height, false); + +// if (removePadding) { +// int padW = (width - imageWidth) / 2; +// int padH = (height - imageHeight) / 2; +// rect.getPath().forEach(point -> { +// keypoints.add(new Point(point.getX() - padW, point.getY() - padH)); +// }); +// rect = +// new Landmark( +// (rect.getX() - padW) / imageWidth, +// (rect.getY() - padH) / imageHeight, +// rect.getWidth() / imageWidth, +// rect.getHeight() / imageHeight,keypoints); +// } else if (applyRatio) { +// rect.getPath().forEach(point -> { +// keypoints.add(new Point(point.getX() / width, point.getY() / height)); +// }); +// rect = +// new Landmark( +// rect.getX() / width, +// rect.getY() / height, +// rect.getWidth() / width, +// rect.getHeight() / height,keypoints); +// } retBB.add(rect); } } diff --git a/face/src/main/java/cn/smartjavaai/face/translator/YoloV8FaceTranslator.java b/face/src/main/java/cn/smartjavaai/face/translator/YoloV8FaceTranslator.java deleted file mode 100644 index 95a3194..0000000 --- a/face/src/main/java/cn/smartjavaai/face/translator/YoloV8FaceTranslator.java +++ /dev/null @@ -1,476 +0,0 @@ -package cn.smartjavaai.face.translator; - -import ai.djl.modality.cv.Image; -import ai.djl.modality.cv.output.*; -import ai.djl.modality.cv.transform.*; -import ai.djl.modality.cv.util.NDImageUtils; -import ai.djl.ndarray.NDArray; -import ai.djl.ndarray.NDList; -import ai.djl.ndarray.types.DataType; -import ai.djl.ndarray.types.Shape; -import ai.djl.translate.*; - -import java.util.*; - -/** - * YoloV8Translator - * @author dwj - */ -public class YoloV8FaceTranslator implements Translator { - - private int maxBoxes; - - private YoloOutputType yoloOutputLayerType; - private float nmsThreshold; - - protected float threshold; -// private BaseImageTranslator.SynsetLoader synsetLoader; - protected List classes; - protected boolean applyRatio; - protected boolean removePadding; - - protected Pipeline pipeline; - private Image.Flag flag; - private Batchifier batchifier; - protected int width; - protected int height; - - - /** - * Constructs an ImageTranslator with the provided builder. - * - * @param builder the data to build with - */ - protected YoloV8FaceTranslator(Builder builder) { - this.yoloOutputLayerType = builder.outputType; - this.nmsThreshold = builder.nmsThreshold; - maxBoxes = builder.maxBox; - this.threshold = builder.threshold; -// this.synsetLoader = builder.synsetLoader; - this.applyRatio = builder.applyRatio; - this.removePadding = builder.removePadding; - this.flag = builder.flag; - this.pipeline = builder.pipeline; - this.batchifier = builder.batchifier; - this.width = builder.width; - this.height = builder.height; - classes = Arrays.asList("face"); - } - - /** - * Creates a builder to build a {@code YoloV8Translator} with specified arguments. - * - * @return a new builder - */ - public static Builder builder() { - return new Builder(); - } - - /** - * Creates a builder to build a {@code YoloV8Translator} with specified arguments. - * - * @param arguments arguments to specify builder options - * @return a new builder - */ - public static Builder builder(Map arguments) { - Builder builder = new Builder(); - builder.configPreProcess(arguments); - builder.configPostProcess(arguments); - return builder; - } - - /** {@inheritDoc} */ - protected DetectedObjects processFromBoxOutput(int imageWidth, int imageHeight, NDList list) { - NDArray rawResult = list.get(0); - NDArray reshapedResult = rawResult.transpose(); - Shape shape = reshapedResult.getShape(); - float[] buf = reshapedResult.toFloatArray(); - int numberRows = Math.toIntExact(shape.get(0)); - int nClasses = Math.toIntExact(shape.get(1)); - int padding = nClasses - classes.size(); - System.out.println(Arrays.toString(reshapedResult.get(0).toFloatArray())); -// if (padding != 0 && padding != 4) { -// throw new IllegalStateException( -// "Expected classes: " + (nClasses - 4) + ", got " + classes.size()); -// } - - ArrayList boxes = new ArrayList<>(); - ArrayList scores = new ArrayList<>(); - ArrayList classIds = new ArrayList<>(); - - // reverse order search in heap; searches through #maxBoxes for optimization when set - for (int i = numberRows - 1; i > numberRows - maxBoxes; --i) { - int index = i * nClasses; - float maxClassProb = buf[index + 4]; -// int maxIndex = -1; -// for (int c = 4; c < nClasses; c++) { -// float classProb = buf[index + c]; -// if (classProb > maxClassProb) { -// maxClassProb = classProb; -// maxIndex = c; -// } -// } -// maxIndex -= padding; - - if (maxClassProb > threshold) { - float xPos = buf[index]; // center x - float yPos = buf[index + 1]; // center y - float w = buf[index + 2]; - float h = buf[index + 3]; - Rectangle rect = - new Rectangle(Math.max(0, xPos - w / 2), Math.max(0, yPos - h / 2), w, h); -// boxes.add(rect); - scores.add(maxClassProb); - classIds.add(0); - List keypoints = new ArrayList<>(); - keypoints.add(new Point(buf[index + 5], buf[index + 6])); - keypoints.add(new Point(buf[index + 8], buf[index + 9])); - keypoints.add(new Point(buf[index + 11], buf[index + 12])); - keypoints.add(new Point(buf[index + 14], buf[index + 15])); - keypoints.add(new Point(buf[index + 17], buf[index + 18])); - Landmark kps = new Landmark(Math.max(0, xPos - w / 2), Math.max(0, yPos - h / 2), w, h, keypoints); - boxes.add(kps); - } - } - - return nms(imageWidth, imageHeight, boxes, classIds, scores); - } - - private DetectedObjects processFromDetectOutput() { - throw new UnsupportedOperationException( - "detect layer output is not supported yet, check correct YoloV5 export format"); - } - - @Override - public DetectedObjects processOutput(TranslatorContext ctx, NDList list) throws Exception { - int imageWidth = (Integer) ctx.getAttachment("width"); - int imageHeight = (Integer) ctx.getAttachment("height"); - switch (yoloOutputLayerType) { - case DETECT: - return processFromDetectOutput(); - case AUTO: - if (list.get(0).getShape().dimension() > 2) { - return processFromDetectOutput(); - } else { - return processFromBoxOutput(imageWidth, imageHeight, list); - } - case BOX: - default: - return processFromBoxOutput(imageWidth, imageHeight, list); - } - } - - @Override - public NDList processInput(TranslatorContext ctx, Image input) throws Exception { - NDArray array = input.toNDArray(ctx.getNDManager(), flag); - NDList list = pipeline.transform(new NDList(array)); - Shape shape = list.get(0).getShape(); - int processedWidth; - int processedHeight; - long[] dim = shape.getShape(); - if (NDImageUtils.isCHW(shape)) { - processedWidth = (int) dim[dim.length - 1]; - processedHeight = (int) dim[dim.length - 2]; - } else { - processedWidth = (int) dim[dim.length - 2]; - processedHeight = (int) dim[dim.length - 3]; - } - ctx.setAttachment("width", input.getWidth()); - ctx.setAttachment("height", input.getHeight()); - ctx.setAttachment("processedWidth", processedWidth); - ctx.setAttachment("processedHeight", processedHeight); - return list; - } - - protected DetectedObjects nms( - int imageWidth, - int imageHeight, - List boxes, - List classIds, - List scores) { - List retClasses = new ArrayList<>(); - List retProbs = new ArrayList<>(); - List retBB = new ArrayList<>(); - - for (int classId = 0; classId < classes.size(); classId++) { - List r = new ArrayList<>(); - List s = new ArrayList<>(); - List map = new ArrayList<>(); - for (int j = 0; j < classIds.size(); ++j) { - if (classIds.get(j) == classId) { - r.add(boxes.get(j)); - s.add(scores.get(j).doubleValue()); - map.add(j); - } - } - if (r.isEmpty()) { - continue; - } - List nms = Rectangle.nms(r, s, nmsThreshold); - for (int index : nms) { - int pos = map.get(index); - int id = classIds.get(pos); - retClasses.add(classes.get(id)); - retProbs.add(scores.get(pos).doubleValue()); -// Rectangle rect = boxes.get(pos); - Landmark rect = boxes.get(pos); - List keypoints = new ArrayList<>(); - if (removePadding) { - int padW = (width - imageWidth) / 2; - int padH = (height - imageHeight) / 2; - rect.getPath().forEach(point -> { - keypoints.add(new Point(point.getX() - padW, point.getY() - padH)); - }); - rect = - new Landmark( - (rect.getX() - padW) / imageWidth, - (rect.getY() - padH) / imageHeight, - rect.getWidth() / imageWidth, - rect.getHeight() / imageHeight,keypoints); - } else if (applyRatio) { - rect.getPath().forEach(point -> { - keypoints.add(new Point(point.getX() / width, point.getY() / height)); - }); - rect = - new Landmark( - rect.getX() / width, - rect.getY() / height, - rect.getWidth() / width, - rect.getHeight() / height,keypoints); - } - retBB.add(rect); - } - } - return new DetectedObjects(retClasses, retProbs, retBB); - } - - public static class Builder { - - private int maxBox = 8400; - - YoloOutputType outputType; - float nmsThreshold; - - protected float threshold = 0.2F; - protected boolean applyRatio; - protected boolean removePadding; - - protected int width = 224; - protected int height = 224; - protected Image.Flag flag; - protected Pipeline pipeline; - protected Batchifier batchifier; - - public Builder() { - this.outputType = YoloOutputType.AUTO; - this.nmsThreshold = 0.4F; - } - - public Builder optOutputType(YoloOutputType outputType) { - this.outputType = outputType; - return this; - } - - public Builder optNmsThreshold(float nmsThreshold) { - this.nmsThreshold = nmsThreshold; - return this; - } - - /** - * Builds the translator. - * - * @return the new translator - */ - public YoloV8FaceTranslator build() { - if (pipeline == null) { - addTransform( - array -> array.transpose(2, 0, 1).toType(DataType.FLOAT32, false).div(255)); - } -// validate(); - return new YoloV8FaceTranslator(this); - } - - protected Builder self() { - return this; - } - - public Builder addTransform(Transform transform) { - if (this.pipeline == null) { - this.pipeline = new Pipeline(); - } - - this.pipeline.add(transform); - return this.self(); - } - - public Builder optApplyRatio(boolean value) { - this.applyRatio = value; - return this.self(); - } - - public Builder optFlag(Image.Flag flag) { - this.flag = flag; - return this.self(); - } - - public Builder setPipeline(Pipeline pipeline) { - this.pipeline = pipeline; - return this.self(); - } - - public Builder setImageSize(int width, int height) { - this.width = width; - this.height = height; - return this.self(); - } - - - public Builder optBatchifier(Batchifier batchifier) { - this.batchifier = batchifier; - return this.self(); - } - - public Builder optThreshold(float threshold) { - this.threshold = threshold; - return this.self(); - } - - /** {@inheritDoc} */ - protected void configPostProcess(Map arguments) { - if (ArgumentsUtil.booleanValue(arguments, "optApplyRatio") || ArgumentsUtil.booleanValue(arguments, "applyRatio")) { - this.optApplyRatio(true); - } - this.threshold = ArgumentsUtil.floatValue(arguments, "threshold", 0.2F); - String centerFit = ArgumentsUtil.stringValue(arguments, "centerFit", "false"); - this.removePadding = "true".equals(centerFit); - String type = ArgumentsUtil.stringValue(arguments, "outputType", "AUTO"); - this.outputType = YoloOutputType.valueOf(type.toUpperCase(Locale.ENGLISH)); - this.nmsThreshold = ArgumentsUtil.floatValue(arguments, "nmsThreshold", 0.4F); - maxBox = ArgumentsUtil.intValue(arguments, "maxBox", 8400); - } - - protected void configPreProcess(Map arguments) { - if (this.pipeline == null) { - this.pipeline = new Pipeline(); - } - - this.width = ArgumentsUtil.intValue(arguments, "width", 224); - this.height = ArgumentsUtil.intValue(arguments, "height", 224); - if (arguments.containsKey("flag")) { - this.flag = Image.Flag.valueOf(arguments.get("flag").toString()); - } - - String pad = ArgumentsUtil.stringValue(arguments, "pad", "false"); - if ("true".equals(pad)) { - this.addTransform(new Pad(0.0)); - } else if (!"false".equals(pad)) { - double padding = Double.parseDouble(pad); - this.addTransform(new Pad(padding)); - } - - String resize = ArgumentsUtil.stringValue(arguments, "resize", "false"); - int w; - int shortEdge; - if ("true".equals(resize)) { - this.addTransform(new Resize(this.width, this.height)); - } else if (!"false".equals(resize)) { - String[] tokens = resize.split("\\s*,\\s*"); - w = (int)Double.parseDouble(tokens[0]); - if (tokens.length > 1) { - shortEdge = (int)Double.parseDouble(tokens[1]); - } else { - shortEdge = w; - } - - Image.Interpolation interpolation; - if (tokens.length > 2) { - interpolation = Image.Interpolation.valueOf(tokens[2]); - } else { - interpolation = Image.Interpolation.BILINEAR; - } - - this.addTransform(new Resize(w, shortEdge, interpolation)); - } - - String resizeShort = ArgumentsUtil.stringValue(arguments, "resizeShort", "false"); - if ("true".equals(resizeShort)) { - w = Math.max(this.width, this.height); - this.addTransform(new ResizeShort(w)); - } else if (!"false".equals(resizeShort)) { - String[] tokens = resizeShort.split("\\s*,\\s*"); - shortEdge = (int)Double.parseDouble(tokens[0]); - int longEdge; - if (tokens.length > 1) { - longEdge = (int)Double.parseDouble(tokens[1]); - } else { - longEdge = -1; - } - - Image.Interpolation interpolation; - if (tokens.length > 2) { - interpolation = Image.Interpolation.valueOf(tokens[2]); - } else { - interpolation = Image.Interpolation.BILINEAR; - } - - this.addTransform(new ResizeShort(shortEdge, longEdge, interpolation)); - } - - if (ArgumentsUtil.booleanValue(arguments, "centerCrop", false)) { - this.addTransform(new CenterCrop(this.width, this.height)); - } - - if (ArgumentsUtil.booleanValue(arguments, "centerFit")) { - this.addTransform(new CenterFit(this.width, this.height)); - } - - if (ArgumentsUtil.booleanValue(arguments, "toTensor", true)) { - this.addTransform(new ToTensor()); - } - - String normalize = ArgumentsUtil.stringValue(arguments, "normalize", "false"); - if ("true".equals(normalize)) { - float[] MEAN = new float[]{0.485F, 0.456F, 0.406F}; - float[] STD = new float[]{0.229F, 0.224F, 0.225F}; - this.addTransform(new Normalize(MEAN, STD)); - } else if (!"false".equals(normalize)) { - String[] tokens = normalize.split("\\s*,\\s*"); - if (tokens.length != 6) { - throw new IllegalArgumentException("Invalid normalize value: " + normalize); - } - - float[] mean = new float[]{Float.parseFloat(tokens[0]), Float.parseFloat(tokens[1]), Float.parseFloat(tokens[2])}; - float[] std = new float[]{Float.parseFloat(tokens[3]), Float.parseFloat(tokens[4]), Float.parseFloat(tokens[5])}; - this.addTransform(new Normalize(mean, std)); - } - - String range = (String)arguments.get("range"); - if ("0,1".equals(range)) { - this.addTransform((a) -> { - return a.div(255.0F); - }); - } else if ("-1,1".equals(range)) { - this.addTransform((a) -> { - return a.div(128.0F).sub(1); - }); - } - - if (arguments.containsKey("batchifier")) { - this.batchifier = Batchifier.fromString((String)arguments.get("batchifier")); - } - - } - } - - public static enum YoloOutputType { - BOX, - DETECT, - AUTO; - - private YoloOutputType() { - } - } - - - -} diff --git a/face/src/main/java/cn/smartjavaai/face/utils/FaceUtils.java b/face/src/main/java/cn/smartjavaai/face/utils/FaceUtils.java index d483b29..8e8d2c4 100644 --- a/face/src/main/java/cn/smartjavaai/face/utils/FaceUtils.java +++ b/face/src/main/java/cn/smartjavaai/face/utils/FaceUtils.java @@ -182,7 +182,12 @@ public class FaceUtils { DetectionRectangle rectangle = detectionInfo.getDetectionRectangle(); graphics.setColor(Color.RED);// 边框颜色 graphics.drawRect(rectangle.getX(), rectangle.getY(), rectangle.getWidth(), rectangle.getHeight()); - drawText(graphics, "face", rectangle.getX(), rectangle.getY(), stroke, 4); + String className = "face"; + if (detectionInfo.getScore() > 0){ + int percent = (int) Math.round(detectionInfo.getScore() * 100); + className = "face " + percent + "%"; + } + drawText(graphics, className , rectangle.getX(), rectangle.getY(), stroke, 4); //绘制人脸关键点 if(detectionInfo.getFaceInfo() != null && detectionInfo.getFaceInfo().getKeyPoints() != null && !detectionInfo.getFaceInfo().getKeyPoints().isEmpty()){ diff --git a/face/src/test/java/Test.java b/face/src/test/java/Test.java index 8462133..5f4c97c 100644 --- a/face/src/test/java/Test.java +++ b/face/src/test/java/Test.java @@ -36,8 +36,10 @@ public class Test { */ public static FaceDetModel getFaceDetModel(){ FaceDetConfig config = new FaceDetConfig(); - config.setModelEnum(FaceDetModelEnum.RETINA_FACE);//人脸检测模型 - config.setConfidenceThreshold(FaceDetectConstant.DEFAULT_CONFIDENCE_THRESHOLD);//只返回相似度大于该值的人脸 +// config.setModelEnum(FaceDetModelEnum.YOLOV8_FACE);//人脸检测模型 + config.setModelPath("/Users/wenjie/Documents/develop/model/yolo-face/yolov8s-face-lindevs.onnx"); +// config.setModelPath("/Users/wenjie/Documents/develop/face_model"); + config.setConfidenceThreshold(0.2);//只返回相似度大于该值的人脸 config.setNmsThresh(FaceDetectConstant.NMS_THRESHOLD);//用于去除重复的人脸框，当两个框的重叠度超过该值时，只保留一个 return FaceDetModelFactory.getInstance().getModel(config); } @@ -81,6 +83,11 @@ public class Test { // } // } // } + + + FaceDetModel faceDetModel = getFaceDetModel(); + R result = faceDetModel.detectAndDraw("/Users/wenjie/Downloads/facetest/00974.png", "/Users/wenjie/Downloads/xx333.png"); + log.info("result:{}", result.isSuccess() + " msg:" + result.getMessage()); }