OpenCV C++ 的 SVM 为相同的类图像返回相同的置信度分数

Question

我已经为 20 个类训练了 SVM，现在对于一个新图像，我想首先将其分类为 20 个训练类之一，然后显示其分类依据的置信度分数。

问题：当我测试第 1 类的一些图像时，它们都返回了相同的置信度分数。我知道每张图片都应该有不同的分数，并以此为基础进行分类。所以我知道我的代码有问题。

我什至把“opencv”框架上的实例变量float *decision_function从protected改成了public并重建了，但是问题还是没有解决。

我想计算正确的分数，因为我想稍后根据置信度分数对与我的查询图像最匹配的图像进行排名。

请告诉我为什么我在同一类中的每张图片都获得相同的置信度分数，以及我在代码中的哪个地方犯了错误。

Ptr<DescriptorMatcher> matcher = DescriptorMatcher::create("FlannBased"); 
Ptr<DescriptorExtractor> extractor = new SurfDescriptorExtractor();
SurfFeatureDetector detector(500);// hemessian threshold

int dictionarySize = 100;  //number of clusters=100
TermCriteria tc(CV_TERMCRIT_ITER, 10, 0.001);
int retries = 1;
int flags = KMEANS_PP_CENTERS;
BOWKMeansTrainer bowTrainer(dictionarySize, tc, retries, flags);
BOWImgDescriptorExtractor bowDE(extractor, matcher);

int main( int argc, char** argv )
{
    int i,j;
    Mat dictionary; // creating a dictionary
    FileStorage fs("dictionary_new.yml", FileStorage::READ);
    fs ["vocabulary"]>> dictionary;
    fs.release();

    bowDE.setVocabulary(dictionary);
    Mat labels(0, 1, CV_32FC1);  //to store labels of images
    Mat trainingData(0, dictionarySize, CV_32FC1);
    int k=0;

    CvSVMParams params;
    params.kernel_type=CvSVM::RBF;
    params.svm_type=CvSVM::C_SVC;
    params.gamma=0.50625000000000009;
    params.C=312.50000000000000;
    params.term_crit=cvTermCriteria(CV_TERMCRIT_ITER,100,0.000001);
    CvSVM svm;

    svm.load("svm_new.yml");
    cout<<"Processing evaluation data..."<<endl;
    string tags[] = {"none","Plane","Bike","Face","Car","Bag","binocular","Gloves","Bread_Maker","Revolver","Ring","Guitar","Elephant","boat","French_horn","Gorilla","Headphone","Flower","Penguin","Tiger","Laptop"};
    Mat evalData(0, dictionarySize, CV_32FC1);
    k=0;
    vector<KeyPoint> keypoint2;
    Mat bowDescriptor2;
    Mat img2; //to load an image

    img2 = imread("3 (14).jpg");
    detector.detect(img2, keypoint2);
    bowDE.compute(img2, keypoint2, bowDescriptor2);
    evalData.push_back(bowDescriptor2);
    float response = svm.predict(bowDescriptor2, true);
    float confidence=0;
    int f_to_i=(int)response;
    confidence=1.0/(1.0+exp(-response));
    cout<<"\n\nIt's a "<<tags[f_to_i]<<" and its confidence: "<<confidence;

    _getch();
    return 0; 
}

计算分数：

1 (1).jpg,Plane,0.731059
1 (2).jpg,Plane,0.731059
1 (3).jpg,Plane,0.731059
1 (4).jpg,Plane,0.731059
1 (5).jpg,Plane,0.731059
1 (6).jpg,Plane,0.731059
1 (7).jpg,Plane,0.731059
1 (8).jpg,Plane,0.731059
1 (9).jpg,Plane,0.731059
1 (10).jpg,Plane,0.731059
1 (11).jpg,Plane,0.731059
1 (12).jpg,Plane,0.731059
1 (13).jpg,Plane,0.731059
1 (14).jpg,Plane,0.731059
1 (15).jpg,Plane,0.731059
1 (16).jpg,Plane,0.731059
1 (17).jpg,Plane,0.731059
1 (18).jpg,Plane,0.731059
1 (19).jpg,Plane,0.731059
1 (20).jpg,Plane,0.731059

Answer 1

我遇到了类似的问题。问题是当您有 2 个以上的类时，您无法使用 svm.predict(bowDescriptor2, true) 获得响应。如果问题是多类 OpenCv SVM 将只返回类标签号。这就解释了为什么你们有相同的价值观。

查看文档：http://docs.opencv.org/modules/ml/doc/support_vector_machines.html#cvsvm-predict

"returnDFVal – 指定返回值的类型。如果为 true 并且问题是 2 类分类，则该方法返回决策函数值，即到边距的有符号距离，否则该函数返回类标签（分类）或估计的函数值（回归）。”

一个可能的解决方案是修改 float CvSVM::predict(const float* row_sample, int row_len, bool returnDFVal) 方法。您可以在那里进行更改以获取每个类的 decisionFunctionValues。

看一段代码： ...

for (i = 0; i < class_count; i++)
{
    for (j = i + 1; j < class_count; j++, df++)
    {
        sum = -df->rho;
        int sv_count = df->sv_count;
        for (k = 0; k < sv_count; k++)
            sum += df->alpha[k] * buffer[df->sv_index[k]];

        vote[sum > 0 ? i : j]++;
    }
}

for (i = 1, k = 0; i < class_count; i++)
{
    if (vote[i] > vote[k])
        k = i;
}
result = returnDFVal && class_count == 2 ? (float)sum : (float)(class_labels->data.i[k]);