Automatic Video Editing Based On Deep Learning

Nowadays,video has been integrated into people's lives,but video processing is still consuming a lot of manpower and material resources.Automatic video editing based on deep learning is the application of face recognition and facial expression recognition technology in the field of computer vision to automatic video processing,which has important research significance and application value.FaceNet is a deep learning framework commonly used in face recognition in recent years.In this paper,by improving the network structure of FaceNet,and designing a De-expression Residue Learning model that integrates self-attention mechanism,automatic video editing of face recognition and expression recognition is realized.The main work of this article is as follows:First,the lightweight FaceNet model based on MobileNet is proposed.On the basis of analyzing the problems existing in FaceNet and the working principle of MobileNet,the lightweight FaceNet model based on MobileNet is proposed.In order to reduce the overall calculation of the network,the model uses deep separable convolutions to decompose standard convolution integrals into deep convolutions and point-by-point convolutions.The model was trained on the CASIA-WebFace and VGGFace2 data sets,and tested on the LFW data set.Experimental results show that the model reduces the network parameters to a large extent while ensuring the accuracy of face recognition,thereby increasing the speed of the system.The model can also perform face recognition on a specific person in the video,and realize automatic video editing based on the one.Then,a face multi-expression recognition model based on improved generative adversarial network is proposed.Based on the analysis of the working principle of the self-attention mechanism,a multi-expression recognition model combining the self-attention mechanism and the De-expression Residue Learning model is proposed.By introducing the self-attention mechanism into the generator of the De-expression Residue Learning model,the accuracy of the generator to generate neutral expressions is improved,which indirectly improves the accuracy of the residual network to classify multiple expressions.The model was trained on the CK+ and Jaffe datasets,and tested on the CK+,Jaffe and MMI data sets.Experimental results show that the model has certain advantages in multi-expression recognition compared with other models.Finally,a prototype system for automatic video editing based on deep learning is designed and implemented.For a given picture,the system can edit a person in the given video that contains the picture automatically,and output a video clip containing only the specific one in the video.In addition,it is also allowed to select a specific expression to output video clips with only specific expressions.The paper has 31 drawings,11 tables and 80 references.