Research And Application Of Portrait Segmentation Based On Attention Mechanism And Guided Filtering

Considering the recent advancements in computer hardware and software,as well as the expansion of the we media sector,there are more and more tasks in image processing,among which portrait segmentation,as an important part,has also become a research hotspot.Since convolutional neural networks have excellent automated learning capabilities,they are frequently employed in portrait segmentation,such as U-Net,Unet++,Attention U-Net.However,the large size difference of the figure in the portrait and the inaccurate edge segmentation bring some difficulties to the portrait segmentation based on convolutional neural networks.Moreover,many portrait segmentation networks now have a large number of parameters and computation,which are not suitable for some scenarios that require lightweight segmentation networks.In this article,in light of a few issues with character images,the semantic segmentation based on convolutional neural network is further researched.The primary research findings consist of:(1)According to the characteristics of portraits,this article extends U-Net network and proposes a portrait segmentation network based on branch attention and dual attention,BADA-UNet.Firstly,a Branch Attention Module(BAM)is introduced into every layer of U-Net network to adjust the size of layers of receptive field,which is caused by the fact that human images have a wide size range and that each layer’s fixed receptive field of neurons.Then,in order to solve the problem that the background of portrait is complex and the background is difficult to distinguish from the edge of figure,this paper introduces a multi-dimensional feature fusion mechanism to obtain more features,and adds a Dual Attention Module(DAM)to simultaneously obtain and integrate the attention information of both channel and space dimensions.Use this information to increase the weight of important features;At last,for alleviating the flaw of large computing consumption and unbalanced positive and negative samples of the divided network,this paper combined the advantages of Binary Cross Entropy Loss in reducing computing resource consumption and Dice Loss in avoiding overfitting,and combined them as the loss function of the network model in this paper.In this paper,two data sets Penn Fudan Ped and CIHP demonstrate its validity.(2)About the flaw of lots of network parameters and floating point calculation,this article proposes a lightweight portrait segmentation network G-UNet.Based on BADA-UNet,it uses depth-separable convolution instead of standard volume and reduces the number of branches of network modules.In addition,considering the effect of guided filtering in preserving portrait edges,this paper adopts the way based on deep learning to improve it and add it to the network,so as to design an end-to-end portrait segmentation network with better integrity.Compared with U-Net,G-UNet has fewer parameters and less computation,and MIo U is even slightly higher.(3)This article,an automatic portrait segmentation system based on G-UNet network model is designed and implemented.The system interface is concise and smooth.It can perform portrait segmentation and background replacement for pictures or videos,and display them.