OpenCV-DNN使用YOLO网络目标检测

开发环境

    VS2013 + OpenCV3.4.1 + Qt5.8.0

实验准备

    yolov2-tiny.cfg

    yolov2-tiny.weights

    或

    yolov2.cfg

    yolov2.weights

    或

    yolov2-voc.cfg

    yolov2-tiny-voc.weights

    以上数据的下载地址戳这里，或者直接下载我的工程(注意：由于*.weights文件过大，无法一并上传到CSDN，*.weights还是需要自行下载，*.weights下载方法如下图)

    

主要代码

加载网络和标签

string modelCfg;
string modelBin;
string classFilename;

if (idx == 0){

	modelCfg = "yolo/yolov2-tiny.cfg";
	modelBin = "yolo/yolov2-tiny.weights";
	classFilename = "yolo/darknet-master/data/coco.names";
}
if (idx == 1){

	modelCfg = "yolo/yolov2.cfg";
	modelBin = "yolo/yolov2.weights";
	classFilename = "yolo/darknet-master/data/coco.names";
}
if (idx == 2){

	modelCfg = "yolo/yolov2-voc.cfg";
	modelBin = "yolo/yolov2-tiny-voc.weights";
	classFilename = "yolo/darknet-master/data/voc.names";
}
try{

	net = dnn::readNetFromDarknet(modelCfg, modelBin);
}catch (cv::Exception &ee){

	QMessageBox::warning(this, "Exception", ee.what());
	if (net.empty()){

		QMessageBox::warning(this, "Exception", "Can't load the network by using the flowing files.");
		return;
	}
}

// Load Label
classNamesVec.clear();
ifstream classNamesFile(classFilename);
if (classNamesFile.is_open()){

	string className = "";
	while (std::getline(classNamesFile, className))
		classNamesVec.push_back(className);
}
else{

	QMessageBox::warning(this, "Exception", "Load label failed.");
	return;
}

前向识别

Mat frame;
image.copyTo(frame);
if (frame.empty()){

	QMessageBox::warning(this, "Warning", "image is empty, please check!");
	return;
}
if (frame.channels() == 4) cvtColor(frame, frame, COLOR_BGRA2BGR);

double ttt = (double)cvGetTickCount();

Mat inputBlob = blobFromImage(frame, 1 / 255.F, Size(416, 416), Scalar(), true, false); //Convert Mat to batch of images
net.setInput(inputBlob, "data");                   //set the network input]
Mat detectionMat = net.forward("detection_out");   //compute output

ttt = (double)cvGetTickCount() - ttt;
ui.labelTime->setText(toChinese("识别时间：") + QString::number(ttt / (cvGetTickFrequency() * 1000000)) + toChinese("秒"));

float confidenceThreshold = ui.dsbConfidence->value();
for (int i = 0; i < detectionMat.rows; i++){

	const int probability_index = 5;
	const int probability_size = detectionMat.cols - probability_index;
	float *prob_array_ptr = &detectionMat.at<float>(i, probability_index);

	size_t objectClass = max_element(prob_array_ptr, prob_array_ptr + probability_size) - prob_array_ptr;
	float confidence = detectionMat.at<float>(i, (int)objectClass + probability_index);

	if (confidence > confidenceThreshold){

		float x_center = detectionMat.at<float>(i, 0) * frame.cols;
		float y_center = detectionMat.at<float>(i, 1) * frame.rows;
		float width = detectionMat.at<float>(i, 2) * frame.cols;
		float height = detectionMat.at<float>(i, 3) * frame.rows;
		Point p1(cvRound(x_center - width / 2), cvRound(y_center - height / 2));
		Point p2(cvRound(x_center + width / 2), cvRound(y_center + height / 2));
		Rect object(p1, p2);

		Scalar object_roi_color(0, 255, 0);

		if (1)//
		{
			rectangle(frame, object, object_roi_color);
		}
		else
		{
			Point p_center(cvRound(x_center), cvRound(y_center));
			line(frame, object.tl(), p_center, object_roi_color, 1);
		}

		String className = objectClass < classNamesVec.size() ? classNamesVec[objectClass] : cv::format("unknown(%d)", objectClass);
		String label = format("%s: %.2f", className.c_str(), confidence);
		int baseLine = 0;
		Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);
		rectangle(frame, Rect(p1, Size(labelSize.width, labelSize.height + baseLine)), object_roi_color, FILLED);
		putText(frame, label, p1 + Point(0, labelSize.height), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 0, 0));
	}
}

Mat detectionMat是一个 输入图像后经过网络前向传播后的输出85*845的结果矩阵，其定义如下图所示：

行向量解析：前面4个元素是用来标记目标在图像上的位置的(被归一化了)，第5个元素是置信概率，值域为[0-1](用来与阈值作比较决定是否标记目标)，后面80个为基于COCO数据集的80分类的标记权重，最大的为输出分类。

列向量解析：为什么是845个？因为YOLO会把图像分成13*13的网格，每个网格预测5个BOX(所以就是13*13*5=845)。每个BOX就是一个行向量(详细的网络细节参考这里)。

实验效果

本地图片

和SSD网络对比

同等条件下和SSD网络比较，识别速度会更快。

附件

    源代码工程戳这里(注：release下的可执行程序可以直接运行，但是要先下载对应的*.weights文件，因为*.weights文件太大，无法一并上传)。