Research On Video Target Detection Algorithm Of Intangible Cultural Heritage Based On Deep Learning

Intangible cultural heritage(hereinafter referred to as "FeiYi")is an important part of China’s traditional culture.Intangible cultural heritage is the core of national culture and the essence of national culture.As a medium of communication,video has its unique advantages,which covers vision and hearing,and is easy to produce and store,so it plays an irreplaceable role in the inheritance,transmission and protection of intangible cultural heritage.However,most of the recorded products of intangible cultural heritage videos are long videos with mixed semantics.Based on the rapid development of short videos and data proof,it can only rely on manual operation,and can not quickly and accurately segment the long video into multiple short videos according to the scene,which has become a major problem hindering the spread of intangible cultural heritage.Therefore,the application of shot boundary detection and boundary frame target detection technology to intangible cultural heritage video will be conducive to the output of intangible cultural heritage short video,and then promote the spread of intangible cultural heritage.Shot boundary detection aims to detect the cut and gradient of the shot in the video,and realize the automatic segmentation of the shot.At present,most shot boundary detection methods often design complex features and similarity measurement methods artificially,and the cost of improving the accuracy is that the algorithm often has high time and space complexity and takes up a lot of computing resources.This paper designs a shot boundary detection model of intangible cultural heritage video based on 3D CNN(Three Dimensional Convolutional Neural Network).The model is divided into two parts,because the artificially designed video frame features are complex,so in the first part,the high-level output of convolutional neural network is used to represent the video frame features,so that the difference between frames can help to abandon the non shot boundary frames that are not in the scope of discussion;in the second stage,the three-demensional convolutional neural network is used to identify the shear in the candidate boundary frames,and the probability value of the network output is used The initial gradient frame is located.The experimental results show that the detection accuracy of this model is 10%higher than that of traditional methods.SSD(single shot multibox detector)learning technology is further studied.The disadvantage of the original SSD algorithm is that the scale of feature map is different when selecting candidate box.To solve this problem,this paper uses RESNET 50 network to replace vgg16 network in the original SSD algorithm,because RESNET 50 network has stronger feature expression ability.At the same time,two modules of feature fusion and hole convolution are designed to combine deep feature map and shallow feature map.Experimental results show that the improved SSD target detection algorithm improves the accuracy of map by 5.2%compared with the original SSD algorithm,and effectively improves the ability of SSD algorithm to detect targets.