使用OpenCV的DNN读取onnx作为预测端实现分类

注意点：
1、size跟训练时一致；
2、训练时用到Normalize的话，因为blobFromImage不支持方差，需要重写blobFromImage；
3、预测结果的格式，确保读取对应指标正确；

class ModelFlipClass
{
private:
	Net net;
public:
	bool Read(string pathNet);
	int Detect(string strImgPath);
};

bool ModelFlipClass::Read(string pathNet)
{
	try {
		net = readNet(pathNet);
	}
	catch (const std::exception&) {
		return false;
	}
	net.setPreferableBackend(cv::dnn::DNN_BACKEND_DEFAULT);
	net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);
	return true;
}

int ModelXRayFlipClass::Detect(string strImgPath)
{
	try 
	{
		Mat imgInput = imread(strImgPath, IMREAD_UNCHANGED);
		Mat imgNorm;
		normalize(imgInput, imgNorm, 0, 255, NORM_MINMAX);//16位转为8位
		convertScaleAbs(imgNorm, imgNorm);
		Mat imRGB;
		cvtColor(imgNorm, imRGB, COLOR_GRAY2BGR);
		const int nHight = 400;//需要跟训练时transforms的Resize保持一致
		const int nWight = 200;
		Mat blob;
		blobFromImage(imRGB, blob, 1 / 255.0, cv::Size(nWight, nHight), Scalar(), true, false);
		net.setInput(blob);
		vector<cv::Mat> netOutputImg;
		net.forward(netOutputImg, net.getUnconnectedOutLayersNames());
		float* pdata = (float*)netOutputImg[0].data;
		vector<float> vecScore;
		for (size_t i = 0; i < 2; i++) //分为两类，所以这里为2
		{
			vecScore.push_back(pdata[i]);
		}
		auto maxPosition = max_element(vecScore.begin(), vecScore.end());
		int nMaxIndex = maxPosition - vecScore.begin();//vdbAmplitude中的值绝对值最大时的索引
		return nMaxIndex;
	}
	catch (const std::exception&) {
		return -1;
	}
}