Trademark Image Retrieval Based On Corner Description And Regional Feature

With the economical development,there is an increasing number of trademark applications and trademark registrations.By the end of 2015,accumulatively there were more than 12 million trademark registrations.In the condition of the enormous trademark database and the strict censorship rules of trademark registration,the research of substituting human review for image retrieval has gained wide attention and an automatic trademark image retrieval system has been one of subjects collectively focused by domestic and overseas academe.This thesis gives a brief analyze of the current trademark retrieval methods and the key techniques in this area.Then a trademark retrieval based on corner descriptor and regional characteristics is proposed,and the experimental results prove the validity and superiority of the proposed algorithm at last.The main contents are as following: 1.A new corner descriptor is presented and applied in image matching.Corner is the key structural point of images,and the neighborhood and distribution of corner describe the main information of images.This thesis presents a new corner descriptor: local statistical distribution(LSD for short)for the local feature of the neighborhood and global statistical distribution(GSD for short)for global distribution of key points.LSD and GSD complement mutually and describe the whole structure of images in combination.Experiments show that the corner feature proposed in this thesis has strong robustness to image scaling,distortion and local noise.2.A two-stage trademark retrieval method based on the combination of corner descriptor and local regional characteristics is proposed.In the first stage,corner matching for the initial screening is based on regional segmentation and cross-matching strategy.Through this method,all the structurally similar images are selected to ensure the recall,while the candidate similar regions are segmented by the corner matching.In the second stage,the corresponding candidate regions are matched by the regional characteristics to ensure the similarity of the image detail and improve the precision.Experiments show that the retrieval framework proposed in this thesis has a good performance on the retrieval of local similarity of trademark images,and has strong robustness to translation,stretching and twisting.