ZF/SmartJavaAI
1
0
Fork 0
mirror of https://github.com/geekwenjie/SmartJavaAI.git synced 2026-09-18 16:39:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

- 【人脸检测】新增6个模型(MTCNN、YOLOV5、RetinaFace小尺寸版)，大幅提升性能

Browse Source 
- 【人脸识别】新增Seetaface6轻量模型
- 【目标检测】支持视频流目标检测（rtsp、视频文件等）
- 【目标检测】支持tensorflow2目标检测模型
- 【目标检测】新增行人检测模型(yolo-person)
- 【通用视觉】新增4个动作识别模型
- 【通用视觉】新增语义分割模型
- 【通用视觉】新增5个实例分割模型(含yolov8-seg、yolov11-seg)
- 【通用视觉】新增yolo-obb11旋转框检测(含yolov11-obb)
- 【通用视觉】新增5个姿态估计模型(含yolov8-pose、yolov11-pose)
This commit is contained in:
dengwenjie
2025-09-07 17:19:19 +08:00
parent 2b044fda29
commit a8e7ce6c4e
102 changed files with 4473 additions and 1368 deletions
Download Patch File Download Diff File Expand all files Collapse all files

65
vision/src/main/java/cn/smartjavaai/action/criteria/ActionRecCriteriaFactory.java
View File

@@ -3,8 +3,10 @@ package cn.smartjavaai.action.criteria;
import ai.djl.Device;
import ai.djl.modality.Classifications;
import ai.djl.modality.cv.Image;
import ai.djl.modality.cv.output.DetectedObjects;
import ai.djl.repository.zoo.Criteria;
import ai.djl.training.util.ProgressBar;
import ai.djl.translate.Translator;
import cn.smartjavaai.action.config.ActionRecModelConfig;
import cn.smartjavaai.action.enums.ActionRecModelEnum;
import cn.smartjavaai.action.model.CommonActionTranslator;
@@ -23,45 +25,48 @@ import java.util.concurrent.ConcurrentHashMap;
public class ActionRecCriteriaFactory {
/**
* 创建动作识别Criteria
* @param config
* @return
*/
public static Criteria<Image, Classifications> createCriteria(ActionRecModelConfig config) {
Device device = null;
if(!Objects.isNull(config.getDevice())){
device = config.getDevice() == DeviceEnum.CPU ? Device.cpu() : Device.gpu(config.getGpuId());
}
Criteria<Image, Classifications> criteria = null;
ConcurrentHashMap params = new ConcurrentHashMap<String, String>();
params.putAll(config.getCustomParams());
if(config.getModelEnum() == ActionRecModelEnum.VIT_BASE_PATCH16_224){
criteria =
Criteria.builder()
.setTypes(Image.class, Classifications.class)
.optModelUrls(StringUtils.isNotBlank(config.getModelPath()) ? null :
config.getModelEnum().getModelUri())
.optModelPath(StringUtils.isNotBlank(config.getModelPath()) ? Paths.get(config.getModelPath()) : null)
.optEngine("PyTorch")
.optDevice(device)
.optProgress(new ProgressBar())
.build();
}else {
Translator<Image, Classifications> translator = getTranslator(config);
if(StringUtils.isBlank(config.getModelEnum().getModelUrl())){
//检查模型路径
if (StringUtils.isBlank(config.getModelPath())){
throw new ActionException("请指定模型路径");
}
int width = 224;
int height = 224;
if(config.getModelEnum() == ActionRecModelEnum.INCEPTIONV3_KINETICS400){
width = 299;
height = 299;
}
criteria =
Criteria.builder()
.setTypes(Image.class, Classifications.class)
.optTranslator(new CommonActionTranslator(width, height))
.optEngine("OnnxRuntime")
.optModelPath(Paths.get(config.getModelPath()))
.optDevice(device)
.optProgress(new ProgressBar())
.build();
}
Criteria<Image, Classifications> criteria =
Criteria.builder()
.setTypes(Image.class, Classifications.class)
.optModelUrls(StringUtils.isNotBlank(config.getModelPath()) ? null : config.getModelEnum().getModelUrl())
.optModelPath(StringUtils.isNotBlank(config.getModelPath()) ? Paths.get(config.getModelPath()) : null)
.optTranslator(translator)
.optDevice(device)
.optProgress(new ProgressBar())
.optEngine(config.getModelEnum().getEngine())
.build();
return criteria;
}
/**
* 获取动作识别Translator
* @param config
* @return
*/
public static Translator<Image, Classifications> getTranslator(ActionRecModelConfig config) {
Translator<Image, Classifications> translator = null;
if(config.getModelEnum() == ActionRecModelEnum.INCEPTIONV1_KINETICS400_ONNX
|| config.getModelEnum() == ActionRecModelEnum.INCEPTIONV3_KINETICS400_ONNX
|| config.getModelEnum() == ActionRecModelEnum.INCEPTIONV3_KINETICS400_ONNX){
translator =new CommonActionTranslator(config.getModelEnum().getInputWidth(), config.getModelEnum().getInputHeight());
}
return translator;
}
}

51
vision/src/main/java/cn/smartjavaai/action/enums/ActionRecModelEnum.java
View File

@@ -6,21 +6,33 @@ package cn.smartjavaai.action.enums;
*/
public enum ActionRecModelEnum {
VIT_BASE_PATCH16_224("djl://ai.djl.pytorch/Human-Action-Recognition-VIT-Base-patch16-224"),
VIT_BASE_PATCH16_224_DJL("PyTorch",0,0,"djl://ai.djl.pytorch/Human-Action-Recognition-VIT-Base-patch16-224"),
INCEPTIONV3_KINETICS400(""),
INCEPTIONV3_KINETICS400_ONNX("OnnxRuntime",299,299,""),
INCEPTIONV1_KINETICS400(""),
INCEPTIONV1_KINETICS400_ONNX("OnnxRuntime",224,224,""),
RESNET18_V1B_KINETICS400(""),
RESNET_V1B_KINETICS400_ONNX("OnnxRuntime",224,224,"");
RESNET34_V1B_KINETICS400(""),
/**
* 模型输入尺寸：宽
*/
private final int inputWidth;
RESNET50_V1B_KINETICS400(""),
/**
* 模型输入尺寸：高
*/
private final int inputHeight;
RESNET101_V1B_KINETICS400(""),
/**
* 模型地址
*/
private final String modelUrl;
RESNET152_V1B_KINETICS400("");
/**
* 模型引擎
*/
private final String engine;
/**
* 根据名称获取枚举 (忽略大小写和下划线变体)
@@ -35,14 +47,27 @@ public enum ActionRecModelEnum {
throw new IllegalArgumentException("未知模型名称: " + name);
}
private final String modelUri;
ActionRecModelEnum(String modelUri) {
this.modelUri = modelUri;
ActionRecModelEnum(String engine, int inputWidth, int inputHeight, String modelUrl) {
this.inputWidth = inputWidth;
this.inputHeight = inputHeight;
this.modelUrl = modelUrl;
this.engine = engine;
}
public String getModelUri() {
return modelUri;
public int getInputWidth() {
return inputWidth;
}
public int getInputHeight() {
return inputHeight;
}
public String getModelUrl() {
return modelUrl;
}
public String getEngine() {
return engine;
}
}

9
vision/src/main/java/cn/smartjavaai/action/model/ActionRecModel.java
View File

@@ -24,15 +24,6 @@ public interface ActionRecModel extends AutoCloseable{
*/
void loadModel(ActionRecModelConfig config);
/**
* 动作检测
* @param base64Image
* @return
*/
default R<Classifications> detectBase64(String base64Image){
throw new UnsupportedOperationException("默认不支持该功能");
}
/**
* 动作检测
* @param image

111
vision/src/main/java/cn/smartjavaai/action/model/ActionRecModelFactory.java Normal file
View File

@@ -0,0 +1,111 @@
package cn.smartjavaai.action.model;
import cn.smartjavaai.action.config.ActionRecModelConfig;
import cn.smartjavaai.action.enums.ActionRecModelEnum;
import cn.smartjavaai.common.config.Config;
import cn.smartjavaai.objectdetection.exception.DetectionException;
import cn.smartjavaai.objectdetection.model.person.CommonPersonDetModel;
import lombok.extern.slf4j.Slf4j;
import java.util.Map;
import java.util.Objects;
import java.util.concurrent.ConcurrentHashMap;
/**
* 动作识别 模型工厂
* @author dwj
*/
@Slf4j
public class ActionRecModelFactory {
// 使用 volatile 和双重检查锁定来确保线程安全的单例模式
private static volatile ActionRecModelFactory instance;
private static final ConcurrentHashMap<ActionRecModelEnum, ActionRecModel> modelMap = new ConcurrentHashMap<>();
/**
* 模型注册表
*/
private static final Map<ActionRecModelEnum, Class<? extends ActionRecModel>> registry =
new ConcurrentHashMap<>();
// 私有构造函数，防止外部创建实例
private ActionRecModelFactory() {}
// 双重检查锁定的单例方法
public static ActionRecModelFactory getInstance() {
if (instance == null) {
synchronized (ActionRecModelFactory.class) {
if (instance == null) {
instance = new ActionRecModelFactory();
}
}
}
return instance;
}
/**
* 获取模型（通过配置）
* @param config
* @return
*/
public ActionRecModel getModel(ActionRecModelConfig config) {
if(Objects.isNull(config) || Objects.isNull(config.getModelEnum())){
throw new DetectionException("未配置模型");
}
return modelMap.computeIfAbsent(config.getModelEnum(), k -> {
return createFaceDetModel(config);
});
}
/**
* 使用ModelConfig创建模型
* @param config
* @return
*/
private ActionRecModel createFaceDetModel(ActionRecModelConfig config) {
Class<?> clazz = registry.get(config.getModelEnum());
if(clazz == null){
throw new DetectionException("Unsupported model");
}
ActionRecModel model = null;
try {
model = (ActionRecModel) clazz.newInstance();
} catch (InstantiationException | IllegalAccessException e) {
throw new DetectionException(e);
}
model.loadModel(config);
return model;
}
/**
* 注册模型
* @param modelEnum
* @param clazz
*/
private static void registerAlgorithm(ActionRecModelEnum modelEnum, Class<? extends ActionRecModel> clazz) {
registry.put(modelEnum, clazz);
}
/**
* 移除缓存的模型
* @param modelEnum
*/
public static void removeFromCache(ActionRecModelEnum modelEnum) {
modelMap.remove(modelEnum);
}
// 初始化默认算法
static {
registerAlgorithm(ActionRecModelEnum.INCEPTIONV1_KINETICS400_ONNX, CommonActionRecModel.class);
registerAlgorithm(ActionRecModelEnum.INCEPTIONV3_KINETICS400_ONNX, CommonActionRecModel.class);
registerAlgorithm(ActionRecModelEnum.RESNET_V1B_KINETICS400_ONNX, CommonActionRecModel.class);
registerAlgorithm(ActionRecModelEnum.VIT_BASE_PATCH16_224_DJL, CommonActionRecModel.class);
log.debug("缓存目录：{}", Config.getCachePath());
}
}

16
vision/src/main/java/cn/smartjavaai/action/model/CommonActionRecModel.java
View File

@@ -68,26 +68,12 @@ public class CommonActionRecModel implements ActionRecModel{
}
}
@Override
public R<Classifications> detectBase64(String base64Image) {
if(StringUtils.isBlank(base64Image)){
return R.fail(R.Status.INVALID_IMAGE);
}
try {
byte[] imageData = Base64ImageUtils.base64ToImage(base64Image);
Image image = ImageFactory.getInstance().fromInputStream(new ByteArrayInputStream(imageData));
return detect(image);
} catch (IOException e) {
throw new DetectionException("读取图片异常", e);
}
}
@Override
public R<Classifications> detect(Image image) {
Classifications classifications = detectCore(image);
// 过滤
if(config.getThreshold() > 0 && CollectionUtils.isNotEmpty(config.getAllowedClasses())
&& Objects.nonNull(classifications) && !classifications.items().isEmpty()){
if(Objects.nonNull(classifications) && !classifications.items().isEmpty()){
classifications = new ClassificationFilter(config.getAllowedClasses(), config.getThreshold()).filter(classifications);
}
return R.ok(classifications);

1
vision/src/main/java/cn/smartjavaai/action/model/CommonActionTranslator.java
View File

@@ -116,7 +116,6 @@ public class CommonActionTranslator implements Translator<Image, Classifications
float[] std = {0.229f * 255, 0.224f * 255, 0.225f * 255};
// 增加 batch 维度，变成 (1, H, W, C)
array = array.expandDims(0);
System.out.println(Arrays.toString(array.getShape().getShape()));
return new NDList(array);
}