Research On Trademark Image Retrieval Based On SVM And Active Learning

With the trend of data explosion nowadays,image data,as a very important data in these data,is applied in various fields.More and more scholars are focusing on the field of image data.Image retrieval is a hot field in the field of image data.Because of the problem of semantic gap between human and computer,the improvement of feedback is introduced.At present,scholars have no longer focused on content based image retrieval research,but put the research direction on the correlation feedback algorithms in image retrieval to further optimize the research.Trademark image,as a symbol of an enterprise,is active in our daily life,and the number of trademarks in China is increasing year by year.But for now,most of trademark image retrieval is only based on the combination of category and text or the content based image retrieval.In trademark image retrieval,the application of relevance feedback algorithm has also become a hot research in trademark image retrieval recently.In view of the application of the correlation feedback algorithm in the field of image retrieval,after reading a large number of SVM active learning related literature,this thesis finds that the application of SVM active learning algorithm in the field of trademark image retrieval can show a good effect.Therefore,a further study on how to achieve efficient application in the field of trademark image retrieval is proposed.In order to reduce the redundancy of samples after each feedback,the improved DBSCAN density clustering algorithm in the density clustering method is used.In order to balance the sample set data,to achieve better accuracy,the oversampling SMOTE algorithm is used.Finally,combining these two methods with SVM active learning algorithm,this thesis proposes an oversampling SVM active learning algorithm based on density clustering.By comparing with the traditional SVM active learning algorithm and the SVM active learning algorithm based on the "V" type deletion method,this paper shows that the algorithm has good performance and has achieved good retrieval effect.