| Portrait segmentation is a more refined form of image segmentation widely used in scenarios such as generating passport photos and editing portraits.Traditional portrait segmentation algorithms have poor performance and heavily rely on manually annotated trimaps and real transparency masks.Additionally,in practical applications,model deployment on back-end servers requires expensive hardware resources and raises privacy concerns.With the advancement of smartphone technology,their computing power has grown stronger,making it possible to perform portrait segmentation on mobile devices.To achieve higher accuracy in portrait segmentation on mobile devices,this study conducted the following research:To improve the quality of portrait segmentation,this paper made some improvements to the MODNet algorithm.By introducing ASPP and combining SE-Block and ASPP to form the SE-ASPP,the SE-EASPP module was finally formed by modifying the SE-ASPP to make it lighter and faster for mobile applications.The SE-EASPP module was used to enhance the semantic feature perception ability of human beings,replacing the original SE-Block.The experimental results show that the improved MODNet algorithm performs better than the original MODNet algorithm.To achieve portrait segmentation on mobile devices,the improved MODNet algorithm model was ported and deployed on an Android phone to implement the function of replacing the background of a portrait.The model was ported to the mobile device using NCNN,completing the core function of replacing the background of a portrait.The portrait background replacement app developed in this study includes functions such as loading photos,selecting backgrounds,replacing backgrounds,and saving images.Experimental tests show that the software can provide portrait background replacement function and can run smoothly on multiple smartphone brands,verifying that the improved MODNet model on mobile devices predicts better than the original MODNet model. |
