\r李 锵，姚麟倩，关 欣\r
\r
AuthorsHTML:\r李 锵，姚麟倩，关 欣\r
\r
AuthorsListE:\rLi Qiang，Yao Linqian，Guan Xin\r
\r
AuthorsHTMLE:\rLi Qiang，Yao Linqian，Guan Xin\r
\r
Unit:\r天津大学微电子学院，天津 300072\r
\r
Unit_EngLish:\rSchool of Microelectronics，Tianjin University，Tianjin 300072，China\r
\r
Abstract_Chinese:\r\r随着深度学习的发展，使用深度卷积神经网络进行关键点定位受到了广泛关注．虽然在人体姿态、人脸识别等多个方面的关键点定位技术已经获得了长足的发展，但是应用于服饰的关键点定位由于其图像背景以及姿态等的多变性依然面临很大的挑战．服饰关键点定位技术在电商以及时尚搭配等方面有很大应用价值，本文将关键点定位应用于时尚领域，提出一种基于级联卷积神经网络的服饰关键点定位算法．该算法的目的是通过级联的两级卷积神经网络，实现对服饰关键点的初步定位以及对困难关键点的定位调整．算法的第\r1\r级以深度残差网络作为特征提取网络，在特征金字塔结构中引入空洞卷积，解决高层特征图感受野大但是空间分辨率低的问题，从而保留更多图像底层细节信息，实现对关键点的初步定位；第\r2\r级将第\r1\r级网络得到的定位结果作为关键点之间的结构先验，结合沙漏网络提取多尺度特征，对困难关键点进行精细调整，进一步提高定位精度．实验选用\r2018 FashionAI \r服饰关键点定位数据集进行训练和测试，将该数据集中对服饰关键点定位的平均归一化误差结果降低到\r3.56\r%\r，充分验证了算法的有效性．与几种常见关键点定位算法进行对比，本文算法在服饰关键点定位任务中取得最好效果，尤其是提高了对困难关键点的定位精度．\r\r
\r
Abstract_English:\r\rWith the development of deep learning\r，\rkey points detection using deep convolutional neural networks\r(\rCNN\r) \rhas attracted extensive attention. Although key points detection for human body posture and face recognition has developed rapidly\r，\rthe application of this technology to clothing faces great challenges because of the variability of the background and posture in clothing pictures. Technology for clothing key points detection has great value for e-commerce and fashion collocation. To apply an algorithm for key point detection to fashion\r，\rin this paper\r，\rwe propose an algorithm for clothing key points detection based on a cascade CNN. This algorithm first detects the key points of clothing and then adjusts difficult key points using a two-level cascade CNN. The first stage of the algorithm detects preliminary key points using ResNet to extract features\r，\rand then\r，\rto retain more detailed image information\r，\ruses dilated convolution to solve the problem of high receptive fields but low spatial resolution in the high-level feature map of a pyramid structure. Using the results from the first stage as a preliminary structure of key points\r，\rthe accuracy is then improved in the second stage by adjusting the difficult key points by combining them with multi-scale features extracted by an hourglass network. We used the 2018 FashionAI clothing landmark dataset for training and testing in the experiment. The normalized error was reduced to 3.56\r%\r in the clothing landmark detection task\r，\rwhich verifies the effectiveness of the network. Compared with the existing algorithm for key points\r，\rthe algorithm proposed in this paper achieves the best result in the task of clothing key points detection\r，\respecially in the detection of difficult key points.\r\r
\r
Keyword_Chinese:级联卷积神经网络；空洞卷积；沙漏网络；关键点定位\r

Keywords_English:cascade convolution neural network；dilated convolutions；hourglass network；key points detection\r


PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6416