Research For Design Patent Image Retrieval Method Based On Co-transduction

With the rapid development of technology,intellectual property rights play a more and more important role in the competition of the country's comprehensive national strength.Patents as an important part of intellectual property,record the inventions and innovations of most inventors.As an important form of inventions in patents,design patent is a kind of new design which include shape,pattern,color and the organic combination of these three,and suitable for industrial applications.Image is the main form of design patent,design patent retrieval based on image retrieval came into being.With the rapid development of Internet and multimedia technology,the number of patent and patent images increased in geometric progression,how to effectively retrieve the information needed by users from the big patent database become the main direction of patent research.Since the introduction of content-based image retrieval in the 1990 s,image retrieval technology has developed rapidly,and content-based design patent retrieval technology development become more and more mature,but still have a large room for improvement in terms of retrieval accuracy and real time search.In order to improve the recall rate and precision rate of image–based design patent retrieval,this paper proposed a new texture feature extraction algorithm for design patent image and proposed a new feature f usion algorithm.This paper carried out research from two aspect which include image features improvement and similarity measure to fuse different feature in another perspective.The main work of the thesis list as follow:(1)Against the object of most design patent image locate in the center of the image,proposed a new method for texture feature extraction-LBP-based on weighted distribution texture entropy.Generate 9 sub-blocks through the non-uniform segmentation of the image,then extract LBP texture feature of the sub-blocks,next statistics the entropy of LBP feature of the sub-blocks,combined the nine sub-blocks' texture information entropy to form a new texture feature.This feature extraction method not only consider the spatial information of the image but also maintain the texture information of the image.(2)Proposed an algorithm of co-training based on re-ranking and SIEAC.Use sample information entropy and confidence(SIEAC)to judge the output samples of co-transduction,and then combine re-ranking idea proposed an algorithm of co-transduction based on re-ranking and SIEAC,and last verified the validity of the two method through corresponding experiments on the design patent dataset.(3)Analyzed five different features,and apply these features to the proposed algorithm in these paper.The five different features are locally binary pattern(LBP),the proposed weighted distribution entropy based LBP(WDE-LBP),histogram of oriented gradients(HOG),and pyramid histogram oriented gradients(PHOG),convolution neural network(CNN).(4)Carried out the verification of the algorithm in the design patent database.In order to improve the accuracy and efficiency of the design patent retrieval,use the co-transduction based on re-ranking and SIEAC algorithm on the design patent image retrieval,and carried out some relevant experiments to prove the advantages of the retrieval performance of the proposed algorithm.