Image Retrieval Method Based On Improved Product Quantization And Its Application In Security Scenarios

In recent years,with the rapid development of computer vision technology,image retrieval as an important research field has also received widespread attention.With the popularization of photoelectric facilities and video surveillance equipment,the demand for image retrieval in security scenarios is increasing.Most security monitoring networks collect multiple monitoring probes,which can generate largescale monitoring video data.Solving the problem of searching target images quickly and accurately in such a large amount of video data has great practical application value.Content based image retrieval technology is a mainstream algorithm in the field of image retrieval,which is roughly divided into two steps: image feature extraction and vector retrieval.Currently,the effective image retrieval algorithms combine feature extraction algorithms based on depth neural networks and vector retrieval algorithms based on product quantization.However,the application of the above algorithms in security scenarios is relatively lacking,and the retrieval accuracy is low when processing images with complex backgrounds.To solve the above problems,the following work has been completed in this article:(1)Pre-processing the surveillance video.Most of the original data obtained in security scenarios are in video format.To retrieve images,it is necessary to first extract and save key frames from the data in video format.The method of extracting frames at equal time intervals is adopted,and the extracted images are classified and stored in chronological order.At the same time,considering the problem of noise in the extracted image,a grayscale transformation normalization algorithm,an adaptive histogram equalization algorithm,and a mean value filter are used for reprocessing.(2)A feature extraction method based on Res Net-50 and channel attention mechanism is proposed.In actual security scenarios,the background of an image captured by the same surveillance camera is basically the same,and the background contains many elements,which are relatively complex.Common feature extraction algorithms focus on the extraction of complex background features,while ignoring the extraction of target features,resulting in too small differences in extracted vector eigenvalues and reduced image retrieval accuracy.To address the above issues,a channel attention mechanism is introduced into the Res Net-50 network,and a triple loss function is used for training to make the algorithm focus more on the target object in the image.The experimental results show that on the Market-1501 dataset,the proposed Res Net-50 feature extraction network combined with attention mechanism improves the m AP value to 85.3% and the Rank-2 value to 87.1%..Compared to other networks under the same training environment,the accuracy of image retrieval has been improved.(3)A vector retrieval method based on improved product quantization is proposed.The product quantization retrieval algorithm includes three steps: Cartesian product decomposition of the original vector space,quantization of the low dimensional vector space,and similarity search.The original product quantization method uses k-means clustering algorithm to quantize low-dimensional vector space,which is affected by the initial point setting and reaches local optimization,resulting in quantization errors.Moreover,many image features have mediations.The clustering of feature vectors using the k-means algorithm belongs to the hardening category,which can cause significant classification errors.Aiming at the above problems,a fuzzy clustering method is proposed to quantify the sub vector space,which avoids the quantization error caused by the hardening score and makes the classification more objective.In the phase of similarity measurement,a weighted asymmetric distance calculation method is used to further reduce the error of distance calculation and improve the accuracy of image retrieval.The advantages of the algorithm are verified through experiments.The experimental results show that on the Oxford 5k dataset,the retrieval accuracy of the FPQ algorithm proposed in this paper is 64.8%,which is 3% higher than the PQ algorithm,2.4% higher than the OPQ algorithm,and 1.3% higher than the IVFPQ algorithm.The average accuracy and average recall rates also have certain improvements,indicating that the FPQ algorithm can outperform the original product quantization and its optimization algorithm.(4)An image retrieval system for security scenarios is designed.Based on security application scenarios,this paper designs an image retrieval system using the above image feature extraction model and vector retrieval model.When a user inputs an image to be queried,all images and related information similar to the image to be queried in the security video database can be retrieved and returned to the user.