Research On Natural Image Matting Algorithm And End-to-End Portrait Matting Algorithm Based On Deep Learning

With the rapid development of social economy and technology,people’s attention to the entertainment field has been increasing,from picture and video compositing to film and TV creation,which have become an indispensable part of daily life.As an important part of image processing technology,image keying technology is widely used in various media productions,such as videos,magazines,advertisements,movies,etc.Traditional matting technology requires expensive and complex equipment,high maintenance costs,and inconvenient deployment,and ordinary people often need to spend a lot of money to rent or build their own green screen matting studio to get a good matting effect.To address this situation,deep learning-based natural image matting algorithms and automatic portrait matting algorithms assist in matting with artificial intelligence image processing technology,helping to simplify the hardware equipment needed for matting,making matting costs plummet.In recent years,research on deep learning-based image matting algorithms has made remarkable breakthroughs with data sets in various scenarios.However,the algorithm development also faces some problems.First,in order to get good results,such algorithms usually need to consume a large amount of computational resources,which makes the deployment cost of the model high;second,in the face of difficult samples such as light bulbs,dandelions and glasses,their imaging quality still needs to be improved.At the same time,automatic matting technology focusing on portraits also faces other problems,for example,the human pose involved in portrait matting varies,and the previous portrait matting algorithms perform poorly in the face of unconventional portraits.To solve the above problems and improve the accuracy of natural image matting,as well as to improve the overall performance of portrait matting in practical applications,this thesis designs a natural image matting algorithm and an end-to-end portrait matting algorithm based on Trimap,respectively.The main work and contributions of this thesis contain the following points:(1)To address the problem that the existing natural image matting algorithm has difficulty in recovering details when facing difficult samples,this thesis improves the existing algorithm by introducing a low-level feature channel guidance module to explore the information implied by these low-level features to help the algorithm recover the transparency mask.Meanwhile,to further improve the overall performance of the model,a dynamic upsampling module is introduced to retain the detail information in the decoder stage.Finally,the algorithm achieves competitive results in terms of both objective evaluation metrics and visual performance.(2)To address the problem of large computer resources required by the portrait matting algorithm,this thesis uses a new process to deal with portrait matting: the whole model is divided into two sub-modules.One is the portrait segmentation module for segmenting portraits and backgrounds,and the other is the detail prediction module for accurately predicting uncertain regions.Dividing the whole task into two independent modules consumes much less resources than doing the regression task globally.(3)To address the problem that the portrait matting algorithm performs poorly in the face of different morphological portraits,this thesis proposes a critical feature extraction module for recovering the details of portraits.Based on the ideas presented in(2),highlevel semantic feature maps are needed to delineate the foreground background and the transparent part of the portrait matting,while more attention needs to be paid to the lowlevel feature maps to predict the transparency of transparent pixels.Therefore,to more fully utilize the information of these two parts of the feature maps,this thesis introduces a critical feature extraction module to further process them.For the feature maps of the intermediate layers,this thesis does not spend a lot of computational resources to process them,so that the performance of the network can be enhanced with almost no additional computational effort introduced.Finally,in order to evaluate the algorithm effect,the proposed natural image matting algorithm based on Trimap and the end-to-end portrait matting algorithm are subjected to validation experiments on their respective datasets,and the experimental results prove that the method proposed in this thesis achieves better results in the same type of matting algorithms.