Research On Lightweighted Robust Video Matting Algorithm

The matting technology is to accurately extract the foreground target area of interest from the picture or video.As the core technology of the special effect synthesis process,matting is widely used in the post-processing process of film and television dramas.Since the matting problem is a typical ill-posed problem,most current matting algorithms solve the matting problem by adding additional constraints.The most commonly used green screen matting method in the film and television industry solves this problem from the perspective of application by limiting the background,but this type of method has relatively high environmental requirements,so it is difficult to popularize.With the development of deep learning technology,the image matting algorithm based on neural network has attracted extensive attention of researchers.Due to the strong generalization ability of neural network,its application scenarios are more extensive.Existing matting algorithms based on neural networks have problems such as large computing power requirements and user interaction,and such methods are still in the development stage.On the basis of image matting,video matting further imposes speed requirements on the algorithm.How to efficiently and accurately extract foreground objects from video sequences has become a research hotspot in recent years.This paper takes people as the research object and based on the existing Robust Video Matting algorithm(Robust Video Matting,RVM),improves the overall network reasoning speed by optimizing the network structure,and proposes the edge details and stability of video matting.New optimization strategy.The main research contents are as follows.(1)Based on the existing advanced robust video matting algorithm,the overall running speed of the network is optimized.This paper first introduces a new backbone network Mobile One-S0 to replace the Mobile Net V3 network,which eliminates the impact of a large number of short connection structures in the backbone network on the speed of model forward propagation reasoning.Improvements have been made to fuse multi-scale features in a more efficient manner,reducing the computational load of this part of the network.Compared with the original model,the improved network model has been improved by 50% in reasoning speed.Although the accuracy is slightly lower than the original network model,it still has obvious advantages compared with other mainstream video matting algorithms.(2)This paper proposes a targeted optimization strategy for the edge details and stability of video matting results.For the edge details of the matting results,this paper proposes an edge supervision loss to enhance the learning of edge features by the network,and combines the idea of hard negative sample mining to eliminate the problem of imbalance between difficult and easy samples of edge pixels.For the stability of video matting results,this paper combines the idea of dropout in the gated recurrent unit of the decoder network,which reduces the network’s dependence on timing features and improves the matting stability of the network in dynamic scenes.