摘要
针对现有深度学习方法中掩模生成质量较低的问题,提出了一种改进的PSPNet掩模优化模型,能够生成较高质量的掩模。保留PSPNet中提取网络ResNet50优秀的残差设计,在此基础上增加卷积注意力机制模块,使模型更加关注掩模边缘,将边缘信息充分的保留至下一层,便于最后上采样生成掩模。上采样过程中只使用双线性插值会导致冗余信息的增加,将双线性插值和像素重组融合,在提高上采样过程的分辨率的同时,保留更多特征,不增加冗余信息,提高掩模生成的质量。最后,加入DICE损失函数,与传统回归损失MSE结合,联合优化模型。结果表明:改进后网络较改进前掩模质量提升了7.1%,同时生成的掩模冗余更少,拐角更加顺滑,便于制造。
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 reorganization can improve the resolution of the upsampling process while retaining more features without adding redundant information and improving the quality of mask generation.Finally,the DICE loss function is added and combined 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.
作者
祁攀
汤府鑫
徐辉
QI Pan;TANG Fuxin;XU Hui(School of Artificial Intelligence,Anhui University of Science and Technology,Huainan Anhui 232001,China;School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)
出处
《兰州工业学院学报》
2024年第1期6-11,共6页
Journal of Lanzhou Institute of Technology
基金
基金委国家重大科研仪器研制项目(62027815)。
