Content Based Video Recognition And Search Algorithm Based On Deep Learning

With the development of information technology and network technology,the application of video has been more and more popular.A large amount of video data is generated on the network each day.Because of the rich content of the large amount of unstructured video data,to effectively manage and retrieve the video data is a long-standing problem.At present,the difficulty of managing and retrieving video data lies in the need for manual intervention in understanding and recognizing video content,which increases the cost of this task.Therefore,intelligent video content recognition and efficient retrieval algorithm are the key technologies to solve this problem.This paper proposes a video content recognition algorithm based on deep learning by studying the characteristics of deep neural network and video content,and introduces neural network to extract the semantic concepts in the video.The main contents and achievements of this paper are as follows:(1)Fusion-SSD network is proposed to extract the semantic concepts of single video frame.In this network,we propose a multi-branch convolution structure,the multi-branch convolution structure improves the network's ability to retain details by introducing parallel structure convolution branches;To enhance the ability of representation of the network,we propose a multi-scale feature fusion algorithm,which effectively resolves the contradiction between object recognition and position regression by fusing the deep and shallow feature.In this paper,we further study the influence of different multi-branch convolution and different fusion methods on the ability to extract network semantic concepts.(2)The Triplet-Stream network is proposed to extract temporal motion semantics from videos.We propose the multi branch feature fusion algorithm and branch weight generation network.The multi feature maps fusion algorithm improves the fusion effect of temporal information and spatial information;the branch weight generation network can adaptively decide the weight of the branch in video action recognition according to the characteristics of video action content,which effectively improves the ability of network to recognize video action.(3)This paper presents a video content search algorithm based on graph structure to represent the output of the neural network.By establishing the relationship between different semantic concepts,the algorithm makes the video content search not limited to the training label data of the neural network,and improves the fuzzy retrieval performance of the video content search.