基於SIFT的特徵匹配

原創

通信程序猿

2020-02-22 00:21

SIFT的全稱是Scale Invariant Feature Transform，尺度不變特徵變換，由加拿大教授David G.Lowe提出的。SIFT特徵對旋轉、尺度縮放、亮度變化等保持不變性，是一種非常穩定的局部特徵。

具體工作原理見SIFT特徵詳解

代碼如下：

# coding=utf-8

import numpy as np
import cv2


sift = cv2.xfeatures2d.SIFT_create()
templet_img = cv2.imread(r'C:\workspace\data\templet.png', 0)
ori_img = cv2.imread(r'C:\workspace\data\object.png', 0)

WIDTH = 512
h, w = ori_img.shape
if h > w:
    object_img = cv2.resize(ori_img, (WIDTH, int(WIDTH * ori_img.shape[0] / ori_img.shape[1])))
else:
    object_img = cv2.resize(ori_img, (int(WIDTH * ori_img.shape[1] / ori_img.shape[0]), WIDTH))

## 角點檢測及特徵提取
kp_templet, des_templet = sift.detectAndCompute(templet_img, None)
kp_object, des_object = sift.detectAndCompute(object_img, None)

## 特徵匹配,利用近似k近鄰算法去尋找一致性
# Brute-Force 方法
bf = cv2.BFMatcher()
matches = bf.knnMatch(des_templet, des_object, k=2)
# FLANN方法(Fast Library for Approximate Nearest Neighbors)
# fla = cv2.FlannBasedMatcher()
# matches = fla.knnMatch(des_templet, des_object, k=2)

## 用比值判別法（ratio test）刪除離羣點
print('match length: ',len(matches))

matchesMask = [[0, 0] for i in range(len(matches))]
obj_point = []
for i, (m1, m2) in enumerate(matches):
    if m1.distance < 0.7 * m2.distance:  # 兩個特徵向量之間的歐氏距離，越小表明匹配度越高。
        matchesMask[i] = [1, 0]
        # pt1 = kp_templet[m1.queryIdx].pt  # queryIdx索引到kp_templet
        pt2 = kp_object[m1.trainIdx].pt  # trainIdx索引到kp_object
        cv2.circle(object_img, (int(pt2[0]), int(pt2[1])), 5, (255, 0, 255), -1)
        obj_point.append(pt2)

# cv2.imshow("object_img", object_img)
# cv2.waitKey(0)

# 顯示匹配結果，匹配點爲藍點, 壞點爲紅點
draw_params = dict(matchColor=(255, 0, 0),
                   singlePointColor=(0, 0, 255),
                   matchesMask=matchesMask,
                   flags=0)

res = cv2.drawMatchesKnn(templet_img, kp_templet, object_img, kp_object, matches, None, **draw_params)
cv2.imshow("Result", res)
cv2.waitKey(0)
cv2.destroyAllWindows()

templet.png 如下：