基于YOLOv5的目标检测——C++版本（展示瑕疵数量、类别，坐标，置信度）

功能：读取文件夹中的所有图片，检测到图片后，保存到另一个文件夹中

先上效果图

先输出上图检测所需时间

下面的4代表瑕疵数量，第一列是瑕疵ID，后面是瑕疵类别、坐标、置信度

代码

主功能

	char class_names[55] = "./defects.names";
	char weights[55] = "./best.torchscript.pt";
	//char source[55] = "./A-M255.bmp";
	string source = "./val";


	//char* class_ = class_names, * weight_ = weights, * source_ = source;
	char* class_ = class_names, * weight_ = weights;

	DETECT detect[100];
	vector<cv::String> image_files;
	glob(source, image_files, false);
	if (!image_files.size()) {
		cout << "No image found!" << endl;
		return 0;
	}
	else {
		for (auto img : image_files) {
			Mat image = imread(img);
			cout << img << endl;
			ModelDete(weight_, class_, detect, image);
		}
	}

图片路径我之前用的是char类型（也可以显示），后来感觉还是string简单一些，循环读取图片，再把图片路径打印出来。

界面

我在此做了个接口，方便后期调用DLL，参数使用char类型。

extern "C" __declspec(dllexport) int ModelDete(char*& torchscript_path, char*& classname_path, DETECT* detect_info, Mat& img_data) {
	string torch = torchscript_path, class_name = classname_path;
	//cout << torch << endl << class_name << endl << img_name << endl;//查看参数

	bool is_gpu = 1;
	torch::DeviceType device_type;
	if (torch::cuda::is_available() && is_gpu) {
		device_type = torch::kCUDA;
		cout << "GPU" << endl;
	}
	else {
		device_type = torch::kCPU;
		cout << "CPU" << endl;
	}

	// load class names from dataset for visualization
	vector<string> class_names = LoadNames(class_name);
	if (class_names.empty()) {
		cout << "Error loading classname!" << endl;
		return -1;
	}

	// load network
	string weights = torch;
	if (weights.empty()) {
		cerr << "Error loading pt!\n";
		return -1;
	}
	auto detector = Detector(weights, device_type);

	//load input image
	//Mat img = imdecode(*img_data,1);
	if (img_data.empty()) {
		cerr << "Error loading the image!\n";
		return -1;
	}
}

算法接口函数使用结构化数组：

struct DETECT {
	int type;//瑕疵类型
	int x;//瑕疵的横坐标
	int y;//瑕疵的纵坐标
	int height;//瑕疵的高度
	int width;//瑕疵的宽度
	double probability;//瑕疵的置信度
};

编写一个函数来保存以下格式：

void pirnt_detect(int num, DETECT* detect_info) {
	string info = to_string(num) + "\t"
		+ to_string(detect_info[num].type) + "\t["
		+ to_string(detect_info[num].x) + "," + to_string(detect_info[num].y) + ','
		+ to_string(detect_info[num].height) + ',' + to_string(detect_info[num].width) + "]\t\t\t"
		+ to_string(detect_info[num].probability) + '\n';
	cout << info;
}

定义另一个函数，然后检测我们的图像：

void DefectsMark(int id, int& cnt, Mat& img,
	const vector<vector<Detection>>& detections,
	const vector<string>& class_names,
	DETECT* detect_info,
	bool label = true) 

if (label) {
		stringstream ss;
		ss << fixed << setprecision(2) << score;
		string s = class_names[class_idx] + " " + ss.str();

		auto font_face = FONT_HERSHEY_DUPLEX;
		auto font_scale = 1.0;
		int thickness = 1;
		int baseline = 0;
		auto s_size = getTextSize(s, font_face, font_scale, thickness, &baseline);
		rectangle(img,
			Point(box.tl().x, box.tl().y - s_size.height - 5),
			Point(box.tl().x + s_size.width, box.tl().y),
			Scalar(0, 0, 255), -1);
		putText(img, s, Point(box.tl().x, box.tl().y - 5),
			font_face, font_scale, Scalar(255, 255, 255), thickness);

		//string str2 = "num:" + std::to_string(num);
	}
	//data save
	if (cnt > 100) {
		cout << "Detect result number is" << num << " > 100, not enough space to save it!\n";
		return;
	}

检测到的图片保存在另一个文件夹中

imwrite("E:/DefectsIdentify/defectsdata/" + to_string(id) + ".bmp", img);

最后展示检测的效果：

图片尺寸4096*4096，太大了就没截全

最近一直在做布检，包括缺陷数据集，也是自己做的。我通常很忙。有需要请留言。