Research And Implementation Of Image Retrieval On Deep Learning

With the explosive growth of emerging media information,it is an essential issue for the scientific researchers how to effectively retrieve sensitive content and meet the needs of law enforcement agencies in guiding Internet public opinion.The retrieval framework with image retrieval technology for deep learning is gradually formed through the analysis and research of sensitive data.The research of this thesis is based on the "Internet Content Supervision Platform LCBIR" project of CETC.The project requires that the sensitive data of video or image from the multimedia content supervision platform can be retrieved,identified,analyzed,and judged effectively,which provides significant clues for other subsystems of the platform on analysis.This thesis focuses on how to use content-based image retrieval(CBIR)technology to effectively retrieve and identify sensitive content data.In CBIR technology,the performance of the retrieval system is directly determined by the image feature extraction method and index retrieval algorithm.On the one hand,in view of the disadvantages of traditional feature extraction algorithm,an improved residual network with the analysis and research of the image feature extraction methods for deep learning is given in this thesis,which can enhance the local invariance and expression ability of extracted features.Additionally,an improvement scheme for image feature extraction by the improved residual network is proposed.On the other hand,in order to meet the needs of large-scale index retrieval,combined with the approximate nearest neighbor search strategy,an improved product quantization algorithm on inverted index is proposed and implemented through the analysis and research of the needs and retrieval strategies of large-scale index retrieval in this thesis.By dividing the feature dimension space,the specified range of the index database is traversed in a non-linear manner to reduce redundant calculation,which achieves the purpose of dynamic addition and deletion of index data,and effectively avoids the high cost of rebuilding the index database in practical applications.Finally,the module performance test and system test are performed on the above implementation.The module test results show that:(1)The improved product quantization algorithm is superior to the MLSH,PQ,HNSW in recall rate,retrieval time and index size.(2)Compared with the existing methods,the method to extract imagefeatures using improved residual network has significantly improved the three performance indicators of CBIR.According to the system test,the CBIR given in this thesis can stably realize requirements of the sensitive content image retrieval system.