Texton Feature Extraction Based Image Retrieval Methods

With the rapid development of Internet and multimedia technology,there has been an explosion of digital images on the Internet.How to search useful images quickly and efficiently from a huge image database(i.e.image retrieval)has been a hot problem in the field of pattern recognition.Traditional image retrieval methods based on text have the problems of ambiguity and time-consuming when annotated by human labor,and is an extremely difficult task to obtain precise images.Content based image retrieval(CBIR)methods extract essential visual features from images to describe image contents,then match features between the query image and images in the database by similarity measure.This method is according with the human visual perception characteristic.Consequently,CBIR is widely utilized.Low level feature extraction of images is the key step in CBIR,which directly affects the results of image retrieval.This thesis studied the texton based feature extraction methods and applied them in image retrieval.On the basis of intense study on color texton and binary texton(local pattern methods)methods,the main contributions of this thesis are summarized as follows:(1)This thesis proposes semicircle local binary patterns structure correlation descriptor(SLBPSCD)to solve the problems of limited texton structure and low feature discrimination.Firstly,a novel semicircle local binary patterns structure texton is defined.Secondly,the structure textons are detected in different color quantization levels.Finally,the spatial distribution and contrast features of new structure texton are extracted.The proposed descriptor obtains much more discriminative structure than conventional color texton methods by taking more structure texton into consideration.Meanwhile,this thesis proposes a novel feature extraction algorithm called three structure descriptor utilizing HSV color space.The proposed method avoids the interference of too much color information and considers information regarding to spatial correlation and information change between structure elements,which improves the discriminative power of feature.Experimental results on different databases and comparisons with conventional color texton methods verify the effectiveness of the two proposed methods in image retrieval.(2)In order to solve the problems of limited local binary pattern and low feature discrimination,this thesis utilizes the directional edge information and spatial structure information of neighborhood pixels in the local structures to improve the discrimination of feature.This thesis introduces Haar feature to local binary pattern and integrates direction feature of neighborhood pixels.This thesis proposes an image feature extraction method of local Haar binary patterns(LHBP)to capture more spatial structure information.Moreover,this thesis extracts texture feature of LHBP and color feature to further improve retrieval precision.Experimental results show that the retrieval precisions of the proposed method reach 78.3%,35.8%,49.5%and 96.3%on the Corel-1000,Corel-5000,Corel-10000 and Coil-100 image databases respectively,which are higher than those of existing methods.(3)In order to improve the feature discriminative power of local pattern methods,this thesis presents some corresponding solutions.Local pattern methods lose the co-occurrence of adjacent local patterns,which causes the loss of a lot of discriminative information.To solve this problem,we firstly propose local binary co-occurrence pattern.It introduces co-occurrence to local binary pattern.In addition,it considers the distribution of every pattern and co-occurrence between adjacent pixels in the image.Local binary co-occurrence pattern improves the discriminative power of feature.Secondly,this thesis proposes co-occurrence local ternary patterns to capture the information of weak edge patterns and relevant information among pixels by combining co-occurrence on local ternary patterns.Finally,experimental results on different databases and comparisons with other methods show that the proposed retrieval methods based on co-occurrence correlation descriptors effectively improve retrieval performance.(4)In order to improve the precision of multi-feature image retrieval,this thesis proposes a hybrid framework for color image retrieval.This hybrid framework fuses the ranking results of bag-of-visual words and intensity-based local difference patterns with color feature(CILDP)by graph fusion method.The proposed fusion method improves the retrieval performance by avoiding problems in non-hybrid methods.Firstly,it avoids the problem that the methods put emphasis on color feature when defining texture directly in color space.In addition,it also avoids another problem that noneffective features may lead to low retrieval performance through combining effective with noneffective features.