摘要
为解决以往算法无法平衡精度和模型大小的问题,提出一种基于上下文学习的轻量级自动抠图算法。采用上下文特征聚合模块和编解码结构相结合的方式进行网络构建,其中编解码器能够有效进行特征提取,通过恢复空间信息捕获更清晰的对象边界;上下文特征聚合模块能够编码多尺度的上下文信息,保留更多细节纹理特征,提高结果的精度。将深度可分离卷积应用到网络主干,形成更快、更强的编解码网络。经Compositon-1k数据集测试,其结果表明,在不进行任何后处理的情况下,该算法精度优于DIM算法,模型参数量较DIM模型降低了98.1%。
To solve the problem that the previous algorithms cannot balance accuracy and model size,a lightweight automatic matting algorithm based on context learning was proposed.The network was constructed by combining the context feature aggregation module and the encode-decoder structure.The encode-decoder effectively extracted features and captured clearer object boundaries by restoring spatial information,while the context feature aggregation module encoded multi-scale context information,retained more detailed texture features,and improved the accuracy of the results.The depthwise separable convolution was applied to the network backbone to form a faster and stronger encoder-decoder network.Tested on the Compositon-1k dataset,the results show that the accuracy of the algorithm is better than the DIM algorithm without any post-processing,and the amount of parameters of the proposed model is 98.1%lower than that of the DIM model.
作者
王文韵
黄根春
田猛
王先培
WANG Wen-yun;HUANG Gen-chun;TIAN Meng;WANG Xian-pei(Electronic Information School,Wuhan University,Wuhan 430072,China)
出处
《计算机工程与设计》
北大核心
2022年第1期94-100,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(51707135)
湖北省重点研发计划基金项目(2020BAB109)
中央高校基本科研业务费专项基金项目(2042019kf1014)。
关键词
深度学习
轻量级
自动抠图
编解码结构
深度可分离卷积
deep learning
lightweight
automatic matting
encode-decoder structure
depthwise separable convolution
