using namespace cv;
    Mat MatBGRImage = imread("wokspace/my_src.jpg");
    auto t0 = iLogger::timestamp_now_float();
    
    Mat RGBImg, ResizeImg;
    cvtColor(MatBGRImage, RGBImg, COLOR_BGR2RGB);
    cv::resize(RGBImg, ResizeImg, Size(224, 224));
    // mean_rgb = [0.485, 0.456, 0.406]
    // std_rgb  = [0.229, 0.224, 0.225]

    int channels = ResizeImg.channels(), height = ResizeImg.rows, width = ResizeImg.cols;

    float* nchwMat = (float*)malloc(channels * height * width * sizeof(float));
    memset(nchwMat, 0, channels * height * width * sizeof(float));

    // Convert HWC to CHW and Normalize
    float mean_rgb[3] = {0.485, 0.456, 0.406};
    float std_rgb[3]  = {0.229, 0.224, 0.225};
    uint8_t* ptMat = ResizeImg.ptr<uint8_t>(0);
    int area = height * width;
    for (int c = 0; c < channels; ++c)
    {
        for (int h = 0; h < height; ++h)
        {
            for (int w = 0; w < width; ++w)
            {
                int srcIdx = c * area + h * width + w;
                int divider = srcIdx / 3;  // 0, 1, 2
                for (int i = 0; i < 3; ++i)
                {
                    nchwMat[divider + i * area] = static_cast<float>((ptMat[srcIdx] * 1.0f/255.0f - mean_rgb[i]) * 1.0f/std_rgb[i] );
                }
            }
        }
    }
    // 复制到GPU
    size_t Src_size = 3 * 224 * 224 * sizeof(float);
    
    for (int i = 0; i < explicit_batch; ++i)
    {
        cudaMemcpy(static_cast<float*>(input_tensor_image->pValue + i * Src_size), 
                nchwMat, Src_size, cudaMemcpyHostToDevice);
    }
    free(nchwMat);
    auto t1 = iLogger::timestamp_now_float();
    auto ms0 = t1 - t0;
    printf("preprocess time: %.3f ms\n", ms0);