Abstract: Aiming at  the problem that the accuracy of feature point matching was low when  image changed, we proposed an improved AKAZE (accelerated-KAZE) algorithm based on perceptual Hash and epipolar constraint. The algorithm  divided feature point matching into two stages: rough matching and fine matching. In the rough matching stage,  ratio of the nearest neighbor and next nearest neighbor of feature points and the perceptual Hash algorithm were used to screen the matching pairs in  the fine matching stage,  the random sample consensus algorithm and epipolar constraint were used to further screen the matching pairs. The simulation results show that, compared with the original algorithm after the random sample consensus algorithm  eliminates the wrong matching pairs, the feature point matching accuracy is still improved by 12.9% on average, and the speed is only 2.4% slower, which can effectively improve the accuracy of matching pairs when the  image changes on the premise of  ensuring the efficiency of the algorithm.

Key words: AKAZE algorithm, feature point matching, nearest neighbor and next nearest neighbor, perceptual Hash, epipolar constraint