Research And Application On Image Retrieval Based On Improved LBP Algorithm

With the arrival of the age of Internet communication and the popularity of electronic devices such as smart phones and networked digital cameras.To meet their needs of searching and acquiring relevant information about the objects they are interested in at any time,users can use their smart devices to search and query in the form of images more and more conveniently.So how to analyze and process image information,how to use image information for relevant retrieval,has become a hot topic of scholars' research,and contentbased image retrieval has been favored by scholars because of its excellent retrieval performance.Based on this,this paper mainly studies the local binary pattern(LBP)which is the extraction method of texture features,and applies it to texture-based image retrieval.The traditional local binary mode(LBP)takes the gray value of the central pixel as the threshold value,and encodes the neighborhood pixel by comparing the gray value and the threshold,so it is not robust to noise.Aiming at the weakness,an adaptive threshold LBP is proposed in this paper.The threshold value is determined by the mean of neighborhood pixels and adaptive parameters,which enhances the anti-noise ability.Aiming at the disadvantage that the original LBP may have different texture features of the same image with different rotation angles due to the problem of image rotation,this paper specifies the starting point of coding,and chooses the direction with the largest gray value of neighborhood pixels as the dominant direction,and combines with the adaptive threshold LBP to form the dominant direction self-adaptive threshold LBP(DSLBP).At the same time,considering the influence of image scale change on texture feature description,different methods of multi-scale analysis of LBP and image pyramid technology are analyzed in detail,and their advantages and disadvantages are pointed out.Considering comprehensively,DSLBP is analyzed at different scales on the Gauss image pyramid.The scale-invariant principal direction self-adaptive threshold LBP is proposed.Firstly,the image is down-sampled with Gauss filter to construct a three-layer image pyramid;secondly,the dominant direction self-adaptive threshold LBP is extracted on each layer of image;finally,the texture feature of the three-layer image is processed according to a certain weight to describe the texture feature of the image.In order to test the performance of the proposed algorithm,image retrieval experiments were carried out on two representative texture databases: KTH-TIPS texture database,UIUC texture database,and it is tested on cloth texture database.The experimental results show that the proposed algorithm can handle noise,rotation,and multi-scale well,it can obtain better retrieval accuracy.Finally,considering the characteristics of clothing images,the algorithm proposed in this paper is applied to the actual application scene of clothing retrieval.The shape features and texture features of clothing are used to describe clothing jointly.