Research On Image Similarity Retrieval Algorithm Based On Perceptual Hashing

In recent years,with the rapid development of multimedia network technology,the way of people's communication has changed from traditional offline communication to online communication.In this context,communication images are rapidly integrated into people's daily life.At the same time,people transmit hundreds of millions of images to cyberspace every day through multimedia and other mean s,waiting for people who need to download and use them.However,the number of images generated is too large,which is not conducive to the user's image retrieval,which puts forward higher requirements for the technology that can quickly and accurately search the image required by the user.Image retrieval based on perceptual hash has become a research hotspot because it can reduce the dimension of image data and improve the retrieval efficiency greatly.Each image can get a hash sequence which can represent the image by reducing the amount of data through feature extraction.By measuring the distance of the hash sequence,we can get the similarity between the images,and then get the retrieval results.However,the selection of image feature extraction method is directly related to the precision and running time of the algorithm.This paper analyzes the problems of traditional perceptual hash algorithm and proposes three improved algorithms.First,in view of the low retrieval accuracy caused by the serious loss of image content information and the difficulty of distinguishing image features in the ahash algorithm,and using the fluctuation degree of a row of data or a column of data can be expressed by variance,an image similar retrieval algorithm based on color variance is proposed.The experimental results show that the average recognition accuracy of the algorithm is 72.17%,the average recognition accuracy of ahash algorithm,phash algorithm and dhash algorithm are 55.33%,85.17% and 74.50%,respective ly.The number of similar images retrieved by the algorithm is significantly more complete than that of ahash algorithm,which is 23.33% higher than that of ahash algorithm.Secondly,the image similarity retrieval algorithm based on double hash is proposed to solve the problem that only one threshold is used to determine the similarity in image similarity retrieval algorithms such as dhash or phash.Experiments show that the average recognition accuracy of this algorithm in image similarity retrieval is 86.17%,while the dhash algorithm is 74.50%,and the phash algorithm is 85.17%.Therefore,the improved algorithm can retrieve all similar images more accurately and completely than the phash / dhash algorithm.Thirdly,the rotation detection experiment proves that the recognition accuracy of the image after rotation is not ideal by using the dhash algorithm.To solve this problem,an improved dhash algorithm against rotation is proposed.Experimental results show that the algorithm can effectively distinguish the image and recognize the image when the image rotates in the inner direction.At the same time,the anti rotation performance of the improved algorithm is better than that of dhash algorithm,and the performance is better than that of the traditional perceptual hash algorithm in the aspect of data attack,such as image contrast adjustment,gamma correction,gaussian filtering,ect.