运行前先下载yolov3的配置文件等，包括：coco.names，yolov3.cfg，yolov3.weights三个文件，可通过wget下载

wget https://github.com/pjreddie/darknet/blob/master/data/coco.names?raw=true -O ./coco.names
wget https://github.com/pjreddie/darknet/blob/master/cfg/yolov3.cfg?raw=true -O ./yolov3.cfg
wget https://pjreddie.com/media/files/yolov3.weights

新建测试类，引入如上三个配置文件。

package com.zk.tobacco.util.yolo;import com.zk.tobacco.util.yolo.modal.BBox;
import org.opencv.core.*;
import org.opencv.dnn.Dnn;
import org.opencv.dnn.Net;
import org.opencv.imgcodecs.Imgcodecs;
import org.opencv.imgproc.Imgproc;import javax.imageio.ImageIO;
import java.awt.image.BufferedImage;
import java.io.*;
import java.util.ArrayList;
import java.util.List;import static org.opencv.highgui.HighGui.namedWindow;
import static org.opencv.imgcodecs.Imgcodecs.imread;
import static org.opencv.imgproc.Imgproc.putText;/*** Copyright (C), 2020-2020* FileName: OpencvYolo3* Author: PJJ* Date: 2020/7/29 12:08* Description: openYolo3*/
public class OpencvYolo3 {static String pro_dir = "D:\\yanchongjiance\\yolo3\\";   // 根路径static float confThreshold = 0.3f;  // 置信度阈值static float nmsThreshold = 0.4f;   // iou阈值static int inpWidth = 416;          // 修改输入图片的宽高static int inpHeight = 416;static List<String> classes = new ArrayList<String>();  // 存放类别的列表static int FONT_HERSHEY_SIMPLEX =1;static int WINDOW_NORMAL=416;public static void main(String[] args) throws Exception{System.loadLibrary(Core.NATIVE_LIBRARY_NAME);//加载配置文件：配置 权重 图片路径 类别文件String modelConfiguration = pro_dir + "yolov3-voc.cfg"; // 模型配置文件String classesFile = pro_dir + "voc.names"; // 模型可识别类别的标签文件String modelWeights = pro_dir + "yolov3.weights"; // 模型权重文件String image_path = "D:\\yanchongjiance\\150\\1.jpg"; //要检测的图片// 进入识别图片的方法detect_image(image_path, modelWeights, modelConfiguration, classesFile);//waitKey(0);}//    检测图片public static void detect_image(String image_path, String modelWeights, String modelConfiguration, String classesFile) throws Exception {// 读取classesFile路径的文件InputStream in = new FileInputStream(classesFile);int iAvail = in.available(); // 适用于本地一次读取多个字节时，返回得到的字节数。byte[] bytes = new byte[iAvail];in.read(bytes);String allContent = new String(bytes); // 文件中的所有内容String[] tempContent = allContent.trim().split("\n"); // allContent去除首尾空格，再按换行符分割。System.out.println(tempContent.length);// 遍历tempContent，添加到保存类别名的列表classes里。for (int i = 0; i < tempContent.length; i++) {classes.add(tempContent[i]);}System.out.println(classes.size());//加载网络配置与训练权重的文件，构建网络Net net =  Dnn.readNetFromDarknet(modelConfiguration, modelWeights);net.setPreferableBackend(Dnn.DNN_BACKEND_OPENCV);net.setPreferableTarget(Dnn.DNN_TARGET_OPENCL);//读入图片，配置宽高Mat im = imread(image_path, Imgcodecs.IMREAD_COLOR); // 读入待检测的图片// 当前是以窗体的形式查看检测后的图片，给窗体命名final String kWinName = "my OpenCV";namedWindow(kWinName, WINDOW_NORMAL);// 将输入图片的宽高重新设置 (416, 416)Mat frame = new Mat();Size sz1 = new Size(im.cols(), im.rows());Imgproc.resize(im, frame, sz1);Mat resized = new Mat();Size sz = new Size(inpWidth, inpHeight);Imgproc.resize(im, resized, sz);float scale = 1.0F / 255.0F; // 像素归一化List<String> ln = net.getLayerNames(); // 获得YOLO各层的名字List<String> x = new ArrayList<String>();List<List<Integer>> out = new ArrayList<List<Integer>>(); // 存放的是列表List<Integer> temp = net.getUnconnectedOutLayers().toList(); // 获得未连接的输出层的索引列表out.add(temp);// out中存放的是一个List ,get(0)得到的就是list i 索引列表List<Integer> i = out.get(0);System.out.println(i.size()); // 3for(int a=0; a<i.size(); a++){String n = ln.get(i.get(a)-1); // 输出层的名字x.add(n); // 找到所有的输出层}ln = x; // 给ln重新赋值System.out.println(ln); // [yolo_82, yolo_94, yolo_106]// 矩阵列表 [Mat[...], Mat[...], Mat[...]]List<Mat> outs = new ArrayList<Mat>();net.forward(outs, ln); // ln此时为输出层的名字列表，向前传播，将得到的检测结果传入outsSystem.out.println(outs);// 进入检测识别方法postprocess_(im, outs);// 检测结束后显示图片//imshow(kWinName, im); // 显示图片// waitKey(300000);}/*** 图片检测识别方法* @param im* @param outs*/public static void postprocess_(Mat im, List<Mat> outs) {System.out.println("检测过程开始");List<Rect> boxes = new ArrayList<Rect>(); // 矩形框列表List<Integer> classIds = new ArrayList<Integer>(); // 类的序号列表List<Float> confidences = new ArrayList<Float>(); // 置信度列表
//        HashMap<String, Object> predicts_dict = new HashMap<String, Object>();// 循环List<Mat>for (int i = 0; i < outs.size(); i++) {Mat mat = outs.get(i);// 循环每一个mat对象// 按行循环for (int j = 0; j < mat.rows(); j++) {int probaility_index = 5; // [x,y,h,w,c,class1,class2] 所以是标号5int size = (int) (mat.cols() * mat.channels());float[] data = new float[size];mat.get(j, 0, data);float confidence = -1;int classId = -1;// 按列循环for (int k = 0; k < mat.cols(); k++) {// 相当于[5:] np.argmax(scores)if (k >= probaility_index && confidence < data[k]) {confidence = data[k]; // 最大值classId = k - probaility_index; // 得到检测得的类别索引}}// 过滤掉置信度较小的检测结果if (confidence > 0.5) {System.out.println("Result  Object:" + j);for (int k = 0; k < mat.cols(); k++) {System.out.println(" " + k + ":" + data[k]);}System.out.println("");float x = data[0]; // centerX 矩形中心点的X坐标float y = data[1]; // centerY 矩形中心点的Y坐标float width = data[2]; // 矩形框的宽float height = data[3]; //矩形框的高float xLeftBottom = (x - width / 2) * im.cols(); // 矩形左下角点的X坐标float yLeftBottom = (y - height / 2) * im.rows(); // 矩形左下角点的Y坐标float xRightTop = (x + width / 2) * im.cols(); // 矩形右上角点的X坐标float yRightTop = (y + height / 2) * im.rows(); // 矩形右上角点的Y坐标// boxes列表填值 Rect对象，参数是两个点boxes.add(new Rect(new Point(xLeftBottom, yLeftBottom), new Point(xRightTop, yRightTop)));confidences.add(confidence);classIds.add(classId);}}}System.out.println(classIds);System.out.println(confidences);System.out.println(boxes.size());System.out.println(boxes);// 非极大值抑制List<Integer> indices = new ArrayList<Integer>();// 此处的非极大值抑制为自己重写的方法。NMSBoxes(boxes, confidences, confThreshold, nmsThreshold, indices);List<Mat> cutImages = new ArrayList<Mat>();Mat cut;int a = 0;// indices : 最终剩下的按置信度由高到低的矩形框序号if (indices.size() > 0) {for (int b = 0; b < indices.size(); b++) {a = a + 1;Rect box = boxes.get(indices.get(b));Point p1 = box.tl(); // 获得左 上角点Point p2 = box.br(); // 获得右下角点int classId = classIds.get(a - 1); // 得到类别序号float confidence = confidences.get(a - 1); // 得到置信度值// 进入画框框方法drawPred_(classId, confidence, im, p1, p2);// 将每个矩形框裁剪cut = im.submat((int)p1.y, (int)p2.y, (int)p1.x, (int)p2.x);cutImages.add(cut);}}System.out.println(cutImages);// 将裁剪后的图片存入本地路径（可省略）String outputFilePath = "D:\\yanchongjiance\\out";for (int i=0; i<cutImages.size(); i++){Mat mat  = cutImages.get(i);MatOfByte mob = new MatOfByte();Imgcodecs.imencode(".jpg", mat, mob);byte[] byteArray = mob.toArray();ByteArrayInputStream in = new ByteArrayInputStream(byteArray);try {BufferedImage image = ImageIO.read(in);OutputStream bOut = new FileOutputStream(outputFilePath + "/" + i + ".jpg");ImageIO.write(image, "jpg", bOut);} catch (IOException e) {e.printStackTrace();}}}//重写的非极大值抑制方法public static void NMSBoxes(List<Rect> boxes, List<Float> confidences, float confThreshold, float nmsThreshold, List<Integer> indices){// 新建一个List 存放BBox的对象List<BBox> bboxes = new ArrayList<BBox>();// 循环向bboxes里添加值：Rect(Point,Point) float confidence, int indexfor(int i=0; i<boxes.size(); i++){BBox bbox = new BBox();bbox.box = boxes.get(i);bbox.confidence = confidences.get(i);bbox.index = i;bboxes.add(bbox);}// 排序算法 根据confidence的属性值，降序排列List bboxesfor (int i = 0; i < bboxes.size(); i++) {for (int j = 0; j < bboxes.size() - i - 1; j++) {if (bboxes.get(j).confidence < bboxes.get(j + 1).confidence) {BBox temp = bboxes.get(j);bboxes.set(j, bboxes.get(j + 1));bboxes.set(j+1, temp);}}}//        for(int i=0; i<bboxes.size(); i++){
//            System.out.println(bboxes.get(i).confidence);
//        }int updated_size = bboxes.size();System.out.println("bboxes.size(): " +bboxes.size());// 循环删除for (int i=0; i<updated_size; i++){// 如果list中的box置信度小于阈值，跳出循环if (bboxes.get(i).confidence < confThreshold){continue;}// indices存放最终剩下的按置信度由高到低的矩形框的序列号indices.add(bboxes.get(i).index);// 比较第一个值（ 最大值）与后面每一个值的交并比// 如果iou>阈值，删除这个框,更改bboxes的长度for (int j=i+1; j<updated_size; j++){float iou = getIouValue(bboxes.get(i).box, bboxes.get(j).box);if(iou > nmsThreshold){bboxes.remove(j);updated_size = bboxes.size();}}}}//获得交并比的方法public static float getIouValue(Rect rect1, Rect rect2){int xx1, yy1, xx2, yy2;xx1 = Math.max(rect1.x, rect2.x);yy1 = Math.max(rect1.y, rect2.y);xx2 = Math.min(rect1.x + rect1.width - 1, rect2.x + rect2.width - 1);yy2 = Math.min(rect1.y + rect1.height -1, rect2.y + rect2.height -1);int insection_width, insection_height;insection_width = Math.max(0, xx2 - xx1 + 1);insection_height = Math.max(0, yy2 - yy1 + 1);float insection_area, union_area, iou;insection_area = insection_width * insection_height;union_area = rect1.width * rect1.height + rect2.width * rect2.height - insection_area;iou = insection_area / union_area;return iou;}//画矩形框，添加文字标签public static void drawPred_(int classId, float confidence, Mat im, Point p1, Point p2){String text;double x = p1.x; // p1 的 x 坐标double y = p1.y; // p1 的 y 坐标// 下面加if语句只是为了区分人和其他类别的不同颜色，改成随机获取颜色也可以if(classId == 0){System.out.println("1");rectangle(im, p1, p2, new Scalar(0, 0, 255), 3);text = classes.get(classId) + ":" + confidence;putText(im, text, new Point(x, y-5), FONT_HERSHEY_SIMPLEX, 0.3, new Scalar(0, 255, 0), 1);}else {System.out.println("2");rectangle(im, p1, p2, new Scalar(0, 255, 0), 3); // 画框text  = String.format("%s %f", classes.get(classId), confidence); // 标签内容System.out.println(text);// 把标签添加到矩形框左上putText(im, text, new Point(x, y-5), FONT_HERSHEY_SIMPLEX, 0.5, new Scalar(0, 0, 255), 1);}}private static void rectangle(Mat im, Point p1, Point p2, Scalar scalar, int i) {}}