计算图像间的类似性能够使用欧氏距离、余弦类似度/做为度量,前者强调点的思想,后者注重线的思想。javascript
欧式距离/Euclidean Distance即n维空间中两个点之间的实际距离。已知两个点A=(a1,a2,...an),B=(b1,b2,...,bn)A=(a1,a2,...an),B=(b1,b2,...,bn),则AB间的距离为:
d(A,B)=[∑(ai−bi)2]−−−−−−−−−−−√(i=1,2,...,n)d(A,B)=[∑(ai−bi)2](i=1,2,...,n)
一样能够利用欧式距离计算图像的类似度,欧式距离越小类似度越大。java
计算欧氏距离:web
double euclidean_distance(Mat baseImg, Mat targetImg)
{
double sumDescriptor = 0;
for (int i = 0; i < baseImg.cols; i++)
{
double numBase = abs(baseImg.at<float>(0, i));
double numTarget = abs(targetImg.at<float>(0, i));
sumDescriptor += pow(numBase-numTarget, 2);
}
double simility = sqrt(sumDescriptor);
return simility;
}
汉明距离/Hamming Distance也能用来计算两个向量的类似度;即经过比较向量每一位是否相同,若不一样则汉明距离加1,这样获得汉明距离。向量类似度越高,对应的汉明距离越小。如10001001和10110001有3位不一样。ide
余弦类似度是利用两个向量之间的夹角的余弦值来衡量两个向量之间的余弦类似度。两个向量越类似夹角越小,余弦值越接近1。
在n维空间中,对于向量A=(a1,a2,...an),B=(b1,b2,...,bn)A=(a1,a2,...an),B=(b1,b2,...,bn),其他弦值为:
cosθ=∑n1(ai×bi)∑n1a2i√×∑n1b2i√cosθ=∑1n(ai×bi)∑1nai2×∑1nbi2svg
double cos_distance(Mat baseImg, Mat targetImg)
{
double squSumB = 0;
double squSumT = 0;
double innerPro = 0;
for (int i = 0; i < baseImg.cols; i++)
{
double numBase = abs(baseImg.at<float>(0, i));
double numTarget = abs(targetImg.at<float>(0, i));
squSumB = squSumB + numBase*numBase;
squSumT = squSumT + numTarget*numTarget;
innerPro = innerPro + numBase*numTarget;
}
double modB = sqrt(squSumB);
double modT = sqrt(squSumT);
double simility = innerPro / (modB*modT);
return simility;
}