Research On Trademark Image Retrieval Based On Region Distribution

With the rapid development of multi-media technology and network technology, the quantity of multi-media information, which is given priority to text, images, voice and video, is growing at an alarming rate in recent years. However, as the complexity of the multimedia information content, the question of how to access and use the vast amounts of network data effectively, becomes an important issue to solve. The trademark image retrieval, a branch of information retrieval technology, is booming in recent years. But as a special trademark search field,the application of trademark image retrieval is relatively less. In view of this situation, this thesis applies the image retrieval method to the trademark retrieval, and puts forward a comprehensive trademark image retrieval algorithm based on regional distribution. The main research content is as follows:1. Aiming at the shortcomings of the conventional methods of image preprocessing, this thesis proposes the improved pretreatment method, to avoid the distortion of the image shape caused by the size normalization towards a great extent.2. The thesis introduces a variety of image segmentation methods, uses the branch cutting method based on the polar coordinate to make up for the inadequacy of quadtree method.3. This thesis proposes a comprehensive trademark image retrieval algorithm based on regional distributiomfirstly, it extractes the eccentricity,7 HU invariant moments and circular degree as global characteristics, does the weighted arithmetic with the set weights, compares the characteristics distance after return, the first retrieval completes; Then it adoptes the branch cutting method which is based on polar coordinates to segment the trademark image, extractes the position of branch cutting image and the pixel density as the local characteristics,the result set of the first retrieve as the image library, obtaines the final retrieval results with the local characteristics match.The thesis compares this algorithm with the other two different methods.(They are in the same framework but with different retrieval characteristics):(1) the trademark image retrieval algorithm based on Hu invariant moments and quadtree (2) the trademark image retrieval algorithm based on the regional characteristics of the sub image. Finally, the thesis statistically analyzes the recall ratio and precision of the retrieval result. The proposed algorithm is verified to has good anti-jamming capability in the retrieval of image translation, rotation, scaling, mirroring and geometric deformation. The retrieve trademark image is consistent approximately with the visual effect judged by human vision system.