Deep Convolutional Network For Shot Video Classification And Recognition System

With the rapid development of modern information technology,huge amounts of short video data have come into view.Algorithms based on deep convolutional neural networks have achieved great success in the field of image recognition,Machines can now train to understand the content of short videos.Some methods extract the input frames into vec-tors,and average or weighting to get the final predict result.They ignor-ing the low-level features between frames.In object detection,current methods use local features to detect objects,ignoring the relations be-tween feature maps.To solve the problems,this paper proposes two algo-rithms:Temporal Part Channel Fusion(TPCF)and Non-Local Feature Fusion.The contributions of this paper summarized as follows:1.Temporal Part Channels Fusion Module is proposed to improve video classification accuracy.TPCF partially fuse the intermediate layer features extracted by the 2D network in the temporal dimension,achieve the accuracy of the 3D network at the cost of the 2D network.2.A Non-Local Feature fusion algorithm is presented.This Non-Local module can be applied to the basic network to fuse the global information between the feature maps,improving the detection performance.3.The overall design and implementation of the short video classifica-tion and recognition system is introduced.The TPCF and Non-Local module are included in the short video classification and recognition sys-tem to improve the overall performance.