Image Retrieval Based On Standard Local Features Combination

With the rapid development of computer and wide application of multimedia, the Contented-Based Image Retrieval has become one of the most active researches in recent days, while most researches are based on traditional global features. The global features extraction techniques are simple and capable to obtain some kind of invariant, but the results are unsatisfactory when images have affine transformation, such as partial occlusion, distortions. However, with the local feature information extracted from the target partial area, it is achievable to indentify targets in the complex backgrounds, so image retrieval based local features has become a hot off the press in the current image retrieval.Due to the diversification of the algorithms in different local feature extraction, it is difficult to use the same retrieval method for all local features, and how to integrate various local features is an urgent issue to solve, so the mechanism of local feature parameters standardization is proposed. According to analysis and study of the existing local features, two different local features MSER and Hessian-Affine, which have different principles and complementary nature, are selected as the underlying features. As the approaches of MSER and Hessian-Affine warping elliptical affine invariant region are different, in order to unify the forms, standard elliptic equation is established and then the axial length and rotation matrix of ellipse with scale are calculated in accordance to the eigenvalues and eigenvector of image feature regions. Finally MSER and Hessian-Affine regions are constructed as a standard elliptical region, and described as a unified parameter form.Based on the local feature standardized mechanism, this thesis constructs a method of image retrieval based on MSER and Hessian-Affine combination. Firstly, this paper achieves a method of the image retrieval based on local features search combination in the same system to avoid the cumbersome establishment of image retrieval systems based on different features. In the system, the appropriate retrieval methods can be adopted according to user's image category: Hessian-Affine is more accurate in image retrieval based full-map and target-based, especially for the corner-rich geometric images; MSER algorithm based on objective adding spatial relation is more accurate to search for images particular for the object with textual information; while the image retrieval based on MHFS can obtain a more satisfactory results for complexity image especially with the structure and the proportion mixed corner point regional information. This search system has greatly enhanced the image retrieval accuracy 11.8% than that of the single feature-based. Secondly, from the perspective of similarity measure, this paper achieves a method of the image retrieval based on MSER and Hessian-Affine weighted combination of the relative scores. By setting the weight to increase probability of the semantically related images retrieved at the same time, the image retrieval accuracy increases by 9.7%. Experimental results show that: from different point of view, which means respective based on the underlying features combination and the principle of similarity measures, the image retrieval can effectively improve the image retrieval accuracy rate taking advantage of combination of two complementary features.