比较两幅图像相似性-基于HOG特征

下面是基于HOG特征的判断两幅图像相似度代码，代码原理很简单，比较两幅图像的HOG特征相关性。

vector<float> extract_hog_features(Mat& image);
float calculate_relation(vector<float>,vector<float>);

int main(int argc,char **argv)
{
    Mat src = cv::imread("./demo.png");
    Mat gray;
    cvtColor(src, gray,COLOR_BGR2GRAY);
    Mat binary;
    threshold(gray, binary, 0, 255, THRESH_OTSU + THRESH_BINARY_INV);
   
    vector<vector<Point>> contours;
    vector<Vec4i> hierarchy;
    findContours(binary, contours, hierarchy, RETR_EXTERNAL, CHAIN_APPROX_SIMPLE,  Point());
    // 按照contours的size大小排序
    for (int i = 0; i < contours.size() - 1; i++)
    {
        for (int j = i + 1; j < contours.size(); j++)
        {
            if (contours[i].size() < contours[j].size())
            {
                swap(contours[i], contours[j]);
            }
        }
    }
    int n = std::min(2, (int)contours.size());
    CV_Assert(n == 2);
    Mat image1,image2;
    
    Rect rect = boundingRect(Mat(contours[0]));
    image1 = src(rect);
    rect = boundingRect(Mat(contours[1]));
    image2 = src(rect);
    vector<float> descriptor1 = extract_hog_features(image1);
    vector<float> descriptor2 = extract_hog_features(image2);
    
    // 计算相关性
    float correlation = calculate_relation(descriptor1, descriptor2);
    std::cout << "correlation is:" << correlation << endl;
    
    return 0;
}

vector<float> extract_hog_features(Mat& image)
{
    Mat gray;
    cvtColor(image, gray, COLOR_BGR2GRAY);
    resize(gray, gray, Size(128, 64));
    HOGDescriptor  hog(Size(128,64),Size(16,16),Size(8,8),Size(8,8),9);
    vector<float> descriptors;
    hog.compute(gray, descriptors);
    
    return descriptors;
}
float calculate_relation(vector<float> descriptor1,vector<float> descriptor2)
{
    CV_Assert(descriptor1.size() == descriptor2.size());
    // 计算内积
    float innerMul = 0.0;
    float amp1 = 0.0;
    float amp2 = 0.0;
    for (int i = 0; i < descriptor1.size(); i++)
    {
        innerMul += descriptor1[i] * descriptor2[i];
        amp1 += pow(descriptor1[i], 2);
        amp2 += pow(descriptor2[i], 2);
    }
    float correlation = innerMul / (sqrt(amp1) * sqrt(amp2));
    
    return correlation;
}

返回结果为：

correlation is:0.891036

相关值在0-1之间，所以据此判断两幅图是高度相关的。

Reference:

https://blog.csdn.net/u010977034/article/details/82733137

https://blog.csdn.net/zhuason/article/details/78933250