基于局部短接单向融合网络的骨架检测  

Skeleton detection based on local short-connection unidirectional fusion networks

作　　者：乔杨 肖士湘 刘悦 焦建彬 QIAO Yang;XIAO Shixiang;LIU Yue;JIAO Jianbin(School of Electronic,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院大学电子电气与通信工程学院,北京100049

出　　处：《中国科学院大学学报（中英文）》2023年第2期250-257,共8页Journal of University of Chinese Academy of Sciences

基　　金：国家自然科学基金(61771447)资助。

摘　　要：近年来,基于侧输出网络的骨架检测方法获得了显著的性能提升。但是,现有方法仍无法解决侧输出结构中高倍上采样和下采样带来的图像失真问题,固定的感受野大小也限制了其视觉特征表达能力。为解决这些问题,提出一种基于侧输出连接的局部短接单向融合网络。该网络由特征提取网络和侧输出网络组成。特征提取网络为深度卷积神经网络,主要用于多层次视觉特征提取。侧输出网络包含局部短接网络和单项融合网络2个模块,其中局部短接网络通过整合感受野邻近特征逐步构建起连续的大感受野特征,而多尺度特征从深到浅的单向融合则实现了对目标从粗糙到精细的刻画。在4种常用骨架检测数据集上的实验结果验证了所提方法的有效性。In recent years,skeleton detection based on side-output network has shown significant effectiveness.However,the existing methods are still unable to tackle the problem of image distortion in side-output structure caused by high-multiplier up/down-sampling,and the fixed receptive field limits the feature expression of the networks.To solve these problems,this paper proposes a local short-connection unidirectional fusion network based on side-output connection,which includes a feature extraction network and a side-output connection network.The feature extraction network is a deep convolutional neural network,which is used for multi-layer feature extraction.The side-output connection network consists of a local short-connection unit and a unidirectional fusion network.The local short-connection gradually constructs the continuous large receptive field features by integrating the adjacent features of receptive field,while the unidirectional fusion of multi-scale features from deep to shallow can achieve the characterization of the object from rough to fine.Experimental results on four commonly used skeleton detection datasets demonstrate the effectiveness of the proposed method.

关 键 词：骨架检测 局部短接单向融合网络 侧输出网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]