Image Retrieval Algorithm Research Base On Saliency Analysis And Multi-feature Fusion

With the dramatic change of Internet technology and the application based on thephoto-sharing, it is more and more difficult for people to find the interesting photos onthe Internet. The traditional image retrieval system based on text has been hard to meetthe demand of accurate retrieval. The content-based image retrieval system emerges andgets considerable development. This thesis focuses on how to make the image retrievalsystems centers on users rather than the computers. The development of this thesis iscomplimented by the two main lines of image saliency analysis and multi-featuresfusion. The main contents are shown as follows:Firstly, starting from the attention mechanisms of the human visual system toexplore the importance and rationality in the image understanding of visual salience.After a comprehensive assessment of the performance and effectiveness of extractionalgorithm of many salient feature points, Harris algorithm is adopted. The detectedsalient points and salient regions provide a basis for the feature extraction of color andtexture respectively.In terms of the color feature extraction, this thesis proposes a method of regioncolor histograms based on the weighted clustering of salient points. Firstly, deeplyanalyzing the differences of a variety of color spaces and applications, adopting the CIEL*a*b*color space from the perspective of unity of color space, and using the K-meansclustering method clusters the color space to a low-dimensional space. Meanwhile, thisarticle puts the K-means clustering to the grouping of salient points. In the process ofclustering, the position and color information of salient points are taken into account,and adding weight value in accordance with the distance of salient points and clustercenter. After that, extracting region color histograms of images and retain the spatialinformation of image content. It has been proved that the method in this thesis has abetter performance than other methods of region color histograms.In terms of the extraction of texture feature, this thesis presents an adaptiveneighbor Local Binary Patterns (LBP) algorithm based on saliency regions. Theneighbor size of traditional LBP algorithm cannot be changed dynamically, which causes many texture calculation errors. To make up for the shortage, this thesis adoptsan improved algorithm, which can obtain the K value by Tamura method used tocalculating the roughness of texture to dynamically change the LBP neighbor size. It isshown that the method in this thesis is superior to the traditional LBP algorithm.In order to further narrow the gap among color, texture feature and high-levelsemantic feature, multi-features fusion method of assign weight is adopted in this article.Meantime, this thesis also introduces the relevant feedback technology based on theweight adjustment. To verify the effectiveness of the method, the author verifies theresults in the public data sets. The experiment results can be seen in this thesis that thealgorithm has a better retrieval performance.