The Research Of Image Retrieval Based On Sal Ient Region And Motif Co-occurrence Matrix Feature

With the rapid growth of the image data as geometric progression,how to achieve a more efficient and accurate retrieval to the image database is the goal and direction of many researchers.Content based image retrieval extract the image color,texture and shape of low-level features for similarity matching,then quickly find the target image.It is an important means in the current information retrieval technology field.The traditional content based image retrieval technique has certain limitations.On the one hand,in order to improve the information description of the main content of the image,many retrieval algorithm filter out the redundant information of the image blindly,and ignore the importance of irrelevant information;On the other hand,the image often contains rich content information.And the use of a single featuredescription can only express a certain aspect of the image information,so it can not explain the integrity of the image description.As an important feature of digital image,color feature is widely used in image feature extraction and indexing.The existing content based image retrieval technology integrate multiple low-level features for image retrieval.Although it avoids the information loss of the original image description and improves the accuracy of image retrieval,it also increases the computational complexity.In order to better describe image content,in this dissertation research,the color features of image is used as a global description,and then combine with salient region color features and motif co-occurrence matrix features by using an effective feature weighted fusion method,respectively.Through describing the feature of the image content from different view,a better retrieval effect is achieved.In this dissertation,we focus on the improvement of the salient region feature and the motif co-occurrence matrix feature,and optimize the image segmentation,feature extraction and feature fusion.The main contents are as follows:1.This dissertation firstly analyzes the relevant technologies in the content based image retrieval field,and a comparative study is made on the feature extraction method of image color,texture and shape features.Then,the image retrieval based on color feature is introduced in detail,and some common methods of similarity measurement and performance evaluation are introduced.Finally,according to the limitation problems of salient region and the motif co-occurrence matrix feature,two kinds of effective image retrieval method is proposed.They are image retrieval method based on global and salient region color feature and image retrieval method based on block motif co-occurrence matrix.2.In order to solve the problem that salient region feature usually ignores the background information,we propose a global and salient region color feature fusion image retrieval method(GASCH).It can emphasis the importance of salient region without losing the background information.Firstly,a quantized HSV color histogram feature is extracted as a global description.Secondly;a salient region detection method is used to separate the salient region from the background region.After that,the color histogram of the salient region is extracted to form a local description.Finally,this dissertation combines these two descriptions by using an adaptive weighting method.The experimental results in Corel 1000 image database show that the proposed method has a better visual effect than the single feature retrieval method,and the retrieval accuracy is improved by at least 9%.3.In order to solve the problem that motif co-occurrence matrix feature does not meet the translation invariance and different sub blocks can be represented by the same motif,we propose a new image retrieval method based on block motif co-occurrence matrix(BMCM).BMCM first divides the image into five regions,and then extracts the quantized HSV color histogram feature,MCM feature and local binary pattern feature from each of the five regions.Considering that different characteristics describes different attributes and content of the image.The method achieves an image retrieval through a weighted fusion of the above three features.Experimental results in Corel 1000 standard image library show that the proposed method has higher precision and obvious advantages in the retrieval of some image categories compared with MCM,BCTF and MCMCM algorithm.