Study on Mask Optimization Algorithm Based on Improved PSPNet
To solve the problem of low mask generation quality in existing deep learning methods,an improved PSPNet mask optimization model is proposed,which can generate higher-quality masks.Retaining the excellent residual design of the extraction network ResNet50 in PSPNet and adding a convolutional block attention module on this basis,the model pays more attention to the edges of the mask and fully retains the edge information to the next layer to facilitate final upsampling to generate the mask.Only using bilinear interpolation in the upsampling process will lead to an increase in redundant information.Combining bilinear interpolation and pixel reorgani-zation can improve the resolution of the upsampling process while retaining more features without adding redun-dant information and improving the quality of mask generation.Finally,the DICE loss function is added and com-bined with the traditional regression loss MSE to optimize the model.The results show that the improved network improves the mask quality by 7.1%compared with the previous improvement.At the same time,the generated mask has less redundancy and smoother corners,making it easier to manufacture.
mask optimizationResNet50convolutional block attention moduleDICE loss
NETL
NSTL
万方数据
