Research Of Visual Object Retrieval Based On Computer Vision

The rapid development of the new generation of IT contributes to the further improvement of new urban infrastructure,and the installation of massive smart cameras for urban governance,transportation,public safety,etc.,which further results in an explosive growth of digital images and video data.Meanwhile,objects for different scale and the details of scene are presented in the high definition view screen.Therefore,how to process massive video and image data efficiently and establish relationships between object and individuals and groups to mine the valid information becoming more and more important.The vision-based object image retrieval technology is vital to solving such problems.As one of the most challenging tasks in the field of computer vision and digital image processing,visual object retrieval mainly aims to construct the object feature database efficiently and retrieve the same or similar object image accurately and quickly from the database.Due to the impact from the data access scale and complex scenario in the actual industrial application environment,an object feature database is huge(usually including more than 100 million pieces of data),and the noise data interference is serious.As a result,it is even more difficult to establish a system of fast and accurate visual object retrieval.In consideration of the current situation,this paper proposes to study the image database building and object retrieval and the innovative work in the following four aspects is specific to visual object retrieval :(1)To alleviate the serious interference of noisy data in the object feature database,object detection is added to the scope of the aforesaid offline research,and effective object data are extracted for the object database.In consideration of the need to process massive digital image data,this paper proposes the concept of EASNet for single stage object detection.While ensuring the real-time detection speed,this network is optimized in aspects,i.e.,the convolution method of the backbone network,the method of information retention in the feature fusion process,and the data imbalance in the loss function.Experimental results show that this network can overcome the problems of multi-scale interference,inaccurate positioning and data imbalance in the real scenes.(2)A multi-level feature fusion algorithm is proposed to accurately return visual object retrieval results and obtain the feature expression in an efficient manner.The bag-of-words model and Hamming Embedding method are adopted for feature fusion and fine-grain quantization.In addition,the regularization diffusion method is adopted to reorder the similarity score results so as to improve the retrieval accuracy.Experiments show that this method can improve the retrieval accuracy effectively.(3)In order to solve the problem of large computational spending and slow return of results from the database retrieval of massive visual objects,this paper designs a new feature index structure,namely Multi-Block N-ary Trie(MBNT).MBNT is used to speed up the retrieval and comparison of object vector features in the Hamming space.The experiment results of algorithm comparison based on the hash table show that MBNT designed for the faster object retrieval can solve the problems of search failure and excessive memory occupation and computational overhead.In addition,when the binary vector features become increasingly compact and the details are better distinguished,the accurate r-nearest neighbor search,compared to the approximate r-nearest neighbor search,has more and more speed advantages in visual object retrieval.(4)A video object detection and retrieval application system is built for the scene of police video combat reconnaissance.The system consists of two subsystems: video analysis and processing system and object retrieval system.The EASNet algorithm for object detection is implemented and used in view resolution subsystem,the multi-level feature fusion algorithm and Multi-Block N-ary Trie(MBNT)algorithm are integrated in the object retrieval subsystem.The system has been tested in many local public security systems and achieved some practical results.