Research On MODNet-based Matting Algorithm For Glasses Images

Image matting is a hot topic in the field of computer vision.Its goal is to perform soft segmentation of images and extract foreground objects in images.This technology has broad application prospects in film and television special effects production,virtual reality,industrial production,etc.For example,in the glasses industry,the traditional glasses try-on is relatively limited,but the target glasses image can be extracted through the image cut-out technology to provide customers with virtual try-on services.In the field of image matting,there is currently a lack of relevant matting algorithms for glasses images,and the existing matting algorithms have low matting accuracy for the glasses image matting task.In view of the above problems,this paper conducts research on the glasses image matting task.On the one hand,for the lack of matting algorithms for glasses images,this paper proposes a glasses image matting algorithm based on MODNet.The algorithm’s network model OMNet constructs a new set of three-branch structures: global semantic segmentation branch,detail matting branch and fusion branch.Each branch internally designs a custom convolution block to optimize the task of eyeglass image matting;more semantic segmentation information of different sizes is transmitted between branches to enhance the interaction between branches.In addition,a pyramid pooling module and multiple attention mechanisms are introduced into the model to improve the influence of important channels and the prediction accuracy of the model.On the other hand,for the lack of glasses image matting dataset,this paper proposes and makes a new glasses image matting dataset GM123.The data set has a total of 25,000 images,including 12,500 glasses images of various materials on the market and their corresponding α true value images.In the process of making the data set,a closed light chamber is built and the glasses image data is collected by a highdefinition camera.In order to verify the effectiveness of the matting algorithm in this paper,the experimental part of this paper first conducts comparative experiments based on public data sets.Then,based on the generated glasses image matting data set,multiple experiments were carried out: In order to verify the prediction accuracy of the matting algorithm in this paper,the effectiveness of the backbone network and each branch,comparative experiments,backbone network experiments and ablation experiments were designed respectively.Finally,the image synthesis experiment is carried out to verify the synthesis effect of the glasses foreground image and the portrait background.The experimental results show that the algorithm proposed in this paper can achieve higher precision,and the evaluation indicators are improved compared with the mainstream matting algorithms.At the same time,the network model OMNet of the algorithm realizes the automatic matting of glasses images,which is easy to deploy and apply.The industrial application of graph technology has certain reference significance.