《计算机应用研究》|Application Research of Computers

基于新的残差沙漏网络的人脸对齐

Face alignment based on new residual hourglass network

免费全文下载 (已被下载 次)  
获取PDF全文
作者 阳邹,邵雄凯,高榕,王春枝
机构 湖北工业大学 计算机学院,武汉 430068
统计 摘要被查看 次,已被下载
文章编号 1001-3695(2020)12-066-3836-05
DOI 10.19734/j.issn.1001-3695.2019.08.0573
摘要 针对传统沙漏网络每一层使用单一感受野来提取特征,缺乏对关键点整体和局部关联信息描述的问题,提出了新的残差沙漏网络(NRHG)方法。该方法通过增加新的卷积分支来增加网络的感受野以更好地提取到不同尺度下的特征信息,同时新的感受野增加了网络对整体信息的描述;针对网络不同层相应调整新增卷积分支感受野大小来平衡feature map分辨率和感受野之间的关系,在更好地保留从局部到整体的结构化信息的同时,突出了网络对局部细节特征信息的描述。最后沙漏网络之间采用中间监督,对每一个沙漏网络输出的结果都进行loss计算,以避免网络深度造成梯度消失的问题。通过在300W、IBUG、COFW数据集上大量的实验证明了该方法的有效性,并且实验结果优于传统的沙漏网络。
关键词 沙漏网络; 感受野; 残差模块; 中间监督
基金项目 国家自然科学基金面上资助项目(61772180)
本文URL http://www.arocmag.com/article/01-2020-12-066.html
英文标题 Face alignment based on new residual hourglass network
作者英文名 Yang Zou, Shao Xiongkai, Gao Rong, Wang Chunzhi
机构英文名 School of Computer Science,Hubei University of Technology,Wuhan 430068,China
英文摘要 For each layer of the traditional hourglass network, using a single receptive field to extract features will lack the description of the overall and local related information of the key points, and it will be difficult to locate in complex situations such as illumination and occlusion. This paper proposed a new residual hourglass network(NRHG), which used the new residual module as the basic unit of the hourglass network. The new residual module added a new convolution branch to increase the network's receptive field for better extraction. To the characteristic information at different scales, the new receptive field increased the description of the network's overall information, and adjusted the size of the new convolutional branch receptive field for different layers of the network to balance the relationship between the feature map resolution and the receptive field. While better retaining the structural information from the local to the whole, it highlighted the local details of the network, and helped to locate key points in complex situations such as illumination and occlusion. In addition, this paper used intermediate supervision between the hourglass networks, and compared the results of each hourglass network output to avoid the problem of gradient degradation caused by network depth. Through a large number of experiments on the 300-W, IBUG, COFW dataset, it proves the effectiveness of the proposed method, and the experimental results are superior to the traditional hourglass network.
英文关键词 hourglass network; receptive field; residual module; intermediate supervision
参考文献 查看稿件参考文献
 
收稿日期 2019/8/9
修回日期 2019/10/24
页码 3836-3840
中图分类号 TP391.41
文献标志码 A