Design And Implementation Of Content-Based Image Retrieval System

With the rapid development of the Internet and imaging technology,the data volume and resolution of digital images have been greatly improved.How to quickly and accurately retrieve target images from massive image data has become an urgent problem to be solved.The content-based image retrieval(CBIR)system shows excellent performance by using the visual information of images for retrieval.Traditional CBIR systems usually use manual methods to obtain low-level visual features of images,and use linear search methods to search databases,resulting in low retrieval accuracy and efficiency.Focusing on the shortcomings and limitations of existing image retrieval methods,the following research has been carried out in this thesis:1)Aiming at the problems of low retrieval efficiency and unbalanced number of positive and negative similar pairs in existing image retrieval,an image retrieval algorithm based on asymmetric deep attention hashing is proposed.This algorithm is based on the Res Net50 network to construct the attention hashing network.On the one hand,the selfattention module and the hashing layer used to generate the hashing codes are introduced into the Res Net50 network to improve the discrimination ability of the hashing codes,and the hashing codes of the query set are obtained through the network;On the other hand,the designed loss function is used to train the network and generate the hashing codes of the database images,which greatly improves the efficiency of hash learning.The experimental results show that the algorithm not only solves the problem of the number imbalance between positive and negative similar pairs,but also effectively improves the efficiency and accuracy of retrieval.2)Aiming at the time-consuming and laborious problem of label labeling in supervised learning and the insufficient semantic category information of original data in existing unsupervised image retrieval,an image retrieval model based on deep triples hashing is proposed.The model firstly uses K-means clustering algorithm to generate image label information in order to construct triplets,and designs a triplet selection strategy to select effective triples.Secondly,the input triplet features are embedded into the potential space by using three variational autoencoders with shared parameters to obtain compact,low-dimensional hashing codes that retain the original data structure information.Finally,a new loss function is designed to encourage the output binary hashing codes to approach the feature representation of the images.Experimental results show that this algorithm effectively improves the discrimination ability of hashing codes and enhances the performance of image retrieval.3)Based on the research results of the above two methods,a CBIR system is designed and developed using Pycharm platform,Tkinter framework and My SQL database.The results show that the designed CBIR system meets the fast and real-time requirements of image retrieval tasks,which shows good retrieval performance and can be applied to multi-scene image retrieval tasks.