Research On Video Retrieval Algorithm Based On Dual Network Model

Internet and social media have been rapidly developed in recent years.Video,as the most effective medium for information dissemination,has appeared widely in people's learning,work,and life.So far,massive amounts of video data have been produced.Therefore,there is an urgent need for a fast and effective video retrieval algorithm for video management and analysis.At the same time,deep learning has made breakthroughs in recent years,and it has performed excellently in the field of image and video processing.This has provided new ideas for video retrieval.This paper combines two kinds of network models in deep learning to study,analyze and improve the key steps in video retrieval: shot segmentation,key frame extraction and feature extraction,and achieve certain results.The main work of this paper is as follows:(1)This paper introduces the research classification and development of video retrieval,and summarizes the ideas,principles,advantages and disadvantages of the classic algorithm of shot segmentation,key frame extraction and feature extraction.(2)This paper proposes a shot segmentation algorithm based on deep learning for the shortcomings of the classic lens segmentation algorithm.This algorithm extracts the depth features of the image and compares the differences between the features as the basis for the shot segmentation.Compared with the classical method,the missed detection and false detection of the shot can be effectively avoided under the scenarios of illumination change,image blur and background jitter.(3)This paper proposes an improved keyframe extraction algorithm based on cluster weights.The algorithm firstly clusters the image frames in the shot,then uses the continuity of the frame serial number to make a second division to obtain the final classification result,then extracts a key frame from each cluster,and finally configure keyframe weights according to the number of image frames in the cluster.Compared with the classical method,this algorithm is more accurate for the key frames extracted by the moving shot,and it also introduces weights to measure the importance,making the number of key frames more reasonable.(4)This article uses a dual network model to construct a video image descriptor.For short shots in video,the algorithm uses convolutional neural networks to extract image features;For the medium and long shots in the video,the convolutional neural network is used to extract the image features while the single-layer long and short memory network is used to extract timing characteristics.Correspondingly,a text-based similarity matching algorithm suitable for this feature is proposed.In the experimental process,the three-part algorithm of shot segmentation,key frame extraction and feature extraction proposed in this paper is concatenated to form a complete video retrieval system.The effect of video retrieval is better than the traditional video retrieval system with SIFT features as the core.In the case where the recall rate is 90%,the precision rate is not lower than 82%.Moreover,with hardware environment acceleration such as GPU,the retrieval time of the system is not slower than the conventional method.