Research And Implementation Of Video Matting Algorithm Based On Transformer

With the sustained development of economy and society and the influence of COVID-19,people have an urgent need for high-quality video background replacement.On the one hand,telecommuting has become the first choice for more people under the epidemic prevention and control,and video background replacement can protect the personal privacy of participants during video conference;On the other hand,the background replacement of short videos can also better meet the entertainment needs of the masses and promote the further development of the short video industry.Video background replacement needs to rely on real-time,efficient and accurate video matting algorithm.At present,the mainstream video matting algorithms generally rely on the auxiliary trimap or background image,and do not pay attention to the extraction of temporal features,so the matting effect is relatively poor.Therefore,it is of great significance to study a more accurate real-time video matting algorithm without auxiliary input.In view of the above application background,relying on the efficient modeling ability of the Transformer architecture for global features,this thesis proposes a video matting algorithm based on Transformer.The main work is as follows:(1)In order to enhance the spatial feature extraction ability,the Transformer module designed in this thesis is used in the corresponding stages of model encoder and decoder.(2)In order to enhance the ability of temporal feature extraction,the model is encouraged to predict the alpha matte at the next time,and the corresponding future frame loss function is used.(3)We design and use hybrid image postprocessing technology to improve the accuracy of matting prediction results.(4)In order to enhance the generalization ability of the algorithm,we collect and make a background data set containing a large number of pictures of different indoor and outdoor scenes.(5)Based on the designed video matting algorithm,this thesis constructs a prototype system of video background replacement application,which can support different modes such as local video background replacement and real-time video stream background replacement.A large number of comparative experiments and the effect of background replacement in practical application scenarios show that the video matting algorithm based on Transformer designed in this thesis can achieve very high matting accuracy in real time and independent of auxiliary input,which verifies the effectiveness of the algorithm proposed in this thesis.