Research On Structured Feature Representation In Images

Managing, classifying, and retrieving multimedia data have broad application valuesand urgent practical needs. Efective image feature representation plays the fundamen-tal role in implementing the above computer vision tasks. However, existing methodsstill sufer from the problems of weak representation ability and discrimination ability.This thesis aims at building an image representation method of high representation anddiscrimination ability, and focuses on the representation method with image structuralfeatures. Motivated by the hierarchical processing mechanism of human vision system(HVS), this thesis proposes a hierarchical structure representation model, and proposesthree structured feature representation methods by considering spatial relationship andscale relationship to characterize the image content, to improve the representation anddiscrimination ability of features. The main work and contributions of this thesis are asfollows.Firstly, motivated by the hierarchical processing mechanisms of HVS, a hierarchicalstructure representation model is proposed. This model utilizes the simple-to-complexhierarchical relation to organize the unordered features to efectively represent the imageinformation. This method addresses the problem of characterizing the image structurefeatures of diferent complexities, and points out two key factors of characterizing theimage structure, scale relationship and spatial relationship. Both the analysis of visualcognition and statistical learning, and the experimental results demonstrate the efective-ness of the proposed model and the importance of scale and spatial relationship in imagerepresentation.Secondly, by introducing the spatial relationship between image features, a spatialneighborhood relationship based single level structure representation method is proposed,to address the problem of low discrimination of other non-structure features. The pro-posed method builds the single level feature pairs by calculating the spatial distribution,describes the pairs by utilizing the relative spatial relationship, and defines the similaritymetric between structure pairs. Furthermore, with this structure representation method, aspatial constraint-based object detection algorithm is proposed. The experimental resultsof object image recognition and Logo detection demonstrate the efectiveness and highdiscriminative ability of the proposed method in image representation. Thirdly, by introducing the scale relationship between image features, a scale rela-tionship based two level structure representation method is proposed, which utilizes themulti-scale feature and scale constraint to improve the discriminative ability of featurerepresentation. The proposed method organizes the unordered features into hierarchicalstructures, characterizes the structures by utilizing the appearance and structure infor-mation, and defines the similarity metric between structure features. A SimHash basedfeature coding method is proposed to achieve a compact and efcient representation. Theexperimental results of object image recognizing, large scale image retrieval and imagematching demonstrate that the proposed feature is more discriminative than the abovestructural features and other feature without structures, and is robust to the variances ofscale and rotation.Finally, based on the single level spatial structure and two level scale structure above,a multi-level structure representation is proposed to describe the content of images. Theproposed method uses a tree structure to organize local multi-scale features based ontheir contained-containing relationship, resulting in further improvement on representa-tion and discrimination ability. A bottom-up coding method is proposed to embed thestructure information into feature vectors. This thesis also proposes a histogram intersectand metric learning based method to calculate the similarity between the structure trees.The experimental results of object image recognizing, texture classification and scene im-age classification demonstrate that the structure tree method is more discriminative thanthe other methods, which demonstrates the efectiveness of the proposed structure treemethod in image representation.In conclusion, through the above-mentioned work, this dissertation makes an inten-sive study on image structural feature representation. The experimental results show that:the structure information of image is important for improving the discriminative ability,and the proposed hierarchical structure representation model is efective for characterizingthe structure information.