嵌入卷积块注意力模块的人像自动抠图算法  

An Automatic Human Matting Algorithm with Convolution Block Attention Module

作　　者：侯玉寒 宋建辉 刘砚菊 刘晓阳 HOU Yuhan;SONG Jianhui;LIU Yanju;LIU Xiaoyang(Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]沈阳理工大学自动化与电气工程学院,沈阳110159

出　　处：《沈阳理工大学学报》2023年第1期35-41,共7页Journal of Shenyang Ligong University

基　　金：辽宁省教育厅高等学校基本科研项目(LJKZ0275);沈阳市中青年科技创新人才支持计划项目(RC210247)。

摘　　要：针对常用人像抠图算法需要输入人工标注三分图及抠图精度不高的问题,提出一种嵌入卷积块注意力模块的人像自动抠图算法。该算法使用三分支网络进行学习:首先预分割分支网络将MobileNetV2与Unet相结合,减少网络参数,引入h-swish激活函数,保留更多有效特征,获取三分图;然后在Alpha抠图分支网络嵌入卷积块注意力模块(CBAM),更好地获取图像多尺度信息,实现Alpha图的初步预测;最后通过细节融合分支网络将以上两个分支的输出进行特征融合,得到Alpha图。实验对比本文算法与现有深度图像抠图(DIM)算法,结果表明,本文算法的绝对误差和(SAD)降低了7.5%、均方误差(MSE)降低了19.4%,实现了人像自动抠图,并获得了良好的抠图效果。Aiming at the problem that the commonly used portrait matting algorithms needs to input manually labeled trisection and the matting accuracy is not high, an automatic human image matting algorithm embedded with convolutional block attention module is proposed.The algorithm uses a three branch network for learning.First, the branch network is pre-split to combine MobileNetV2 with Unet, network parameters are reduced, h-swish activation function is introduced, more effective features are retained, and a tripartite graph is obtained;Then, the convolutional block attention module(CBAM)is embedded in the Alpha matting branch network to better obtain the image multi-scale information and realize the preliminary prediction of the Alpha image;Finally, the outputs of the above two branches are fused by the detail fusion branch network to obtain the Alpha graph.The experimental results show that the sum of absolute differences(SAD)is reduced by 7.5%,and the mean squared error(MSE)is reduced by 19.4% compared with the existing deep image matting(DIM)algorithm, automatic human image matting is realized, and good matting effect is obtained.

关 键 词：深度学习 人像自动抠图 注意力模块 特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]