Key Technology Research On Large-scale Image Target Retrieval Based On Product-RSOM

The research about large-scale retrieval had drawn much attention in the image processing field during last few decades.However,because of the fast-growing database and imperfect algorithms,the retrieval methods we have cannot meet people's requirements such as retrieval speed and memory cost.In order to improve user experience on large-scale image retrieval,some key technologies have been proposed in this paper,which got impressive results in the retrieval system we built.The main works of this paper could be summarized as following aspects:(1)A compressed coding algorithm based on Product-RSOMFirst,we design a new method,Product-RSOM,on the base of RSOM[1],which aiming at solving the problem of high-dimension vectors and massive storage spaces.This method improves the efficiency of feature matching through dividing and quantizing SIFT descriptors in block.In this way,we could utilize sub-SIFT to achieve the process of mutual matching which,on a certain extent,reduces the complexity of algorithm for the multiple cuts of dimension.Meanwhile,the product-quantization of high-dimension samples also means the compressing of storage spaces which could employ numbered quantization centers to approximate all the training samples in database if the quantization error could be controlled within permissible range.Experimental results proved that when the number of leafnodes on 16D-RSOM tree come to 60000,the quantization error among training samples could be less than 0.5 percent.(2)The realization of image retrieval with SPGCIn order to avoid the unexpected situation during the process of feature detection and region matching,this paper brings in the function of SPGC(Similarity Propagation based Graph Clustering)to fulfill the sharing of information and improve the outcomes of retrieval.For example,if the region of interesting is vague on a certain image,we could still figure it out with another similar image.The main works of this algorithm,on the base of Product-RSOM,include the division,quantization,coding and reconstructing,and finally complete the process with RANSAC[2] precision matching and inverted index[5].Experimental results showed that this algorithm's rate of recognition could reach 95% and the index time could be less than 15 millisecond over the database containing 50 thousand images.(3)the construction of large-scale image retrieval systemIn order to verify the feasibility of algorithm,we build a million-scale images retrieval system based on Product-RSOM.It mainly consists of two modules,one for database training and another for image index.In database training module,we could accomplish the training of vast samples,the recording of key information and also the propagating of mutual similarities.As for image recognition module,both of accuracy-first searching and breadth-first searching could be achieved.Besides,this system has the function of incremental training [3] and similarity propagation [4] which earn an more impressive performance in the aspect of man-machine interaction.This system has been in service for a technology company for months and has got the positive reply from company staff.