用于人体解析的多尺度边缘指导网络  

Multi-scale Edge-guided Network for Human Parsing

作　　者：廖勤俭 李东[1] 

机构地区：[1]广东工业大学自动化学院,广东广州510006

出　　处：《工业控制计算机》2023年第7期67-69,共3页Industrial Control Computer

基　　金：广东省自然科学基金资助项目(2021A1515011867)。

摘　　要：人体解析是语义分割的一种,属于非常有挑战性的密集预测任务,其目标是对人体各部位及服饰等进行象素级别的分类。目前的人体解析算法通常会使用边缘检测作为辅助任务来提升人体解析性能。但是现有的方法中对于边缘特征的提取和利用能力较差,限制了其辅助监督能力。因此,基于注意力机制设计了多尺度边缘特征解码器以及边缘指导融合模块来分别提升边缘特征的提取和利用能力。此外将所提出的方法应用于Transformer骨干网络,提供了一个人体解析新基线。最后,在Pascal-person-part和LIP数据集上进行了充分实验,证明了该方法的有效性并达到了先进水平。Human parsing is a type of semantic segmentation,which is a very challenging intensive prediction task with the goal of classifying various class of the human body,clothing,etc.at the pixel level.Existing human parsing algorithms usually use edge detection as an auxiliary task to improve segmentation accuracy.However,the existing methods'performance for edge feature extraction and utilization is poor,which limits its auxiliary supervision capability.Therefore,this paper designs a multi-scale edge feature decoder and an edge guidance module based on the attention mechanism to enhance the extraction and utilization of edge features,respectively,and applies the proposed method to the Transformer backbone network and provide a new human parsing baseline.Finally,this paper conducts sufficient experiments on Pascal-person-part and LIP dataset to demonstrate the effectiveness of our methods and to achieve advanced level.

关 键 词：人体解析 语义分割 边缘检测 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]