摘要
自然拍摄的人体照片由于背景图案较为复杂,采用传统基于图片色彩空间或能量梯度的图像处理方法难以准确地识别人体的轮廓。采用神经网络的方法,可以提高识别的精度。但是,一般的神经网络方法由于计算量与参数规模较大,难以在移动终端部署。因此,提出了一种轻量级的神经网络策略以提取人体轮廓。该网络采用MobileNet V2与U-Net框架,通过构建特定姿态的人体数据集进行训练,识别相应的人体轮廓形状。人体轮廓经过提取关键点、拟合回归分析等后续处理,可估算人体的尺寸。该方法可应用在移动终端上,通过拍摄的人体照片的方法测量人体的尺寸。实验表明,该方法能准确地提取复杂背景照片中的人体轮廓并测量尺寸,在速度与存储占用方面较一般神经网络有一定优势。
Due to the complexity of background,it is difficult to accurately recognize the contour of human body by traditional image-processing methods based on color space or energy gradient.Neural network can improve the accuracy of recognition.However,due to the large scale of computations and parameters,it is difficult to deploy the general neural network methods in mobile devices.Therefore,we proposed a lightweight neural network to extract human body contours.This network utilized MobileNet V2 and U-Net framework to recognize the contours of human bodies by building a human-body dataset with specific poses for training.The contours of human bodies can be used to measure the sizes of human bodies after the subsequent processes,such as the extraction of key points and analysis of fitting regression.This method can be applied to mobile terminals to measure the body sizes by taking pictures.Experiments show that this method can accurately extract the contours of human bodies in photos with complex backgrounds and measure the body sizes,and that it possesses some advantages over the general neural network in terms of speed and storage.
作者
吴泽斌
张东亮
李基拓
麻菁
信玉峰
WU Ze-bin;ZHANG Dong-liang;LI Ji-tuo;MA Jing;XIN Yu-feng(College of Computer Science and Technology,Zhejiang University,Hangzhou Zhejiang 310027,China;School of Mechanical Engineering,Zhejiang University,Hangzhou Zhejiang 310027,China;School of Art and Design,Sanming University,Sanming Fujiang 365004,China)
出处
《图学学报》
CSCD
北大核心
2020年第5期740-749,共10页
Journal of Graphics
基金
国家自然科学基金项目(61732015,61972340)
浙江省重点研发计划项目(2018C01090)。
关键词
图像处理
轮廓提取
人体尺寸测量
轻量级神经网络
深度学习
image processing
contour extraction
body measurement
lightweight neural network
deep learning
