从一组图像中查找图像匹配矩阵

Question

我有以下代码，其中使用距离阈值和David Lowe的比率测试，使用cv2 orb查找图像之间的匹配项。同一组图像（a，b）的匹配与a-> b和b <-a不同，因此我将两个匹配都保留下来，只保留对称匹配。现在，我有一组93张图像，并且必须在图像之间创建一个距离矩阵（其中距离不是对称匹配的数目）。这种矩阵计算需要花费大量时间，如何提高速度？

 import cv2
 import numpy as np
 import random
 im1 = cv2.imread("eiffel_18.jpg",0)
 im2 = cv2.imread("eiffel_19.jpg",0)
 im3 = cv2.imread("bigben_3.jpg",0)
 im4 = cv2.imread("bigben_2.jpg",0)
 im5 = cv2.imread("bigben_12.jpg",0)
 im6 = cv2.imread("notredame_1.jpg")

 img_list=[im1,im2,im3,im4,im5]
 jk=['a','b','c','d','e']
 no_images=len(img_list)






def no_matches(img1,img2):
    # detect features 
     orb = cv2.ORB_create(nfeatures=500)
     (keypoints1, descriptors1) = orb.detectAndCompute(img1, None)
    (keypoints2, descriptors2) = orb.detectAndCompute(img2, None)

distances={}
for i in range(0, len(keypoints1)):
    temp=[]
    dist=0
    for j in range(0,len(keypoints2)):
       distance=cv2.norm( descriptors1[i], descriptors2[j], cv2.NORM_HAMMING)
       dist+=distance
       temp.append((j,distance))
       if(j==len(keypoints2)-1):
           t=sorted(temp,key=lambda x: x[1])
           num=t[0]
           den=t[1]
           distances[i]=(num,den,dist)

pairs=[]

## Ratio test and thresholding according to D.Lowe's paper 
first_count=0
second_count=0
for keys,values in distances.items():
    numerator=values[0][1]
    denominator=values[1][1]
    t=values[2]/500




    if(numerator<64):
        first_count+=1
        if(numerator/float(denominator))<0.85:
            second_count+=1

            end1 = tuple(np.round(keypoints1[keys].pt).astype(int))

            end3 = tuple(np.round(keypoints2[values[0][0]].pt).astype(int))
            pairs.append((end1,end3))


return pairs,second_count,first_count

 ############################ distt is a matrix with all the distances between 
 all images ####################################
 ############################ Will be used to look up inter image distances in 
 K-means #######################################

 distt=np.zeros((no_images,no_images),dtype=np.float32)


 distt.fill(-1)
 print("before",distt)

 ##ignoring distances of images with themselves
 for i in range(distt.shape[0]):
 print(i)
 for j in range(distt.shape[1]):
    if(i!=j):
        if(distt[i][j]==-1):
            relation,s1,f1=no_matches(img_list[i],img_list[j])
            relation_symmetric,s2,f2=no_matches(img_list[j],img_list[i])
            ctr=0
            for match1 in relation:
                for match2 in relation_symmetric:
                    if(match1[0]==match2[1] and match1[1]==match2[0]):
                        ctr+=1
            total_matches=(s1+s2)+0.001
            if(min(s1,s2)>0.8*max(s1,s2)):
                total_matches=0.5*total_matches
            distt[i][j]=((ctr)/total_matches)*100
            distt[j][i]=((ctr)/total_matches)*100

 print(distt)