关键词: 复杂背景, 深度学习, 人体轮廓提取, 整体嵌套边缘检测网络, 二维照片

Abstract:

Objective For the collection of human body images in non-contact two-dimensional measurement systems, most of the early studies reduced the difficulty of human contour extraction by means of restricting the shooting background. In order to acquire human contour and parameter information quickly and easily from photographs with a complex background, this study proposes the use of holistically-nested edge detection (HED) deep learning model to achieve the extraction of human contours and perform parameter extraction analysis.

Method A training set of 43 200 images was established by creating a human contour label map, data enhancement and pre-processing with 450 human photos with different backgrounds as the original image dataset. A deep learning network model was used for training and learning, and the optimal edge detection model was constructed after repeated training and tuning to achieve automatic human contour extraction. In addition, in order to verify that the extracted human silhouette form is consistent with the real human form, 13 human parameters are selected for error analysis.

Results After training and tuning the deep learning model to obtain the optimal model for human contour extraction, the test-set images were subsequently output (Fig.5), which shows that the optimized HED network model training extracteds clear human contours without other cluttered background information, but the human contour edge lines was thick. Therefore, the Zhang-Suen edge refinement algorithm was adoptrd to refine the human contours (Fig.5(d)). It can be seen that after the edge refinement, not only the edges of the human contour can be detailed and clear with no break in continuity, but also some non-human contour details in the original contour map can be removed to make the background cleaner. In addition, in order to verify the accuracy of extracting human contours, 40 subjects were selected and subjected to 3-D anthropometric measurements and 2-D photographs, respectively. Based on the human body photograph to extract the human body contour and feature point identification positioning, 13 measurements were extracted and compared with the corresponding 3-D point cloud manual measurements (Tab.3). It can be seen that the error range of 3 angle parameters are between 0.125 3°and 1.862 2°, and the error range of 10 ratio parameters are between 0.000 2 and 0.081 7. The data indicated that there was no significant difference between the contour measurements and the 3-D real values, and a high consistency was maintained between them, which further verified the feasibility and accuracy of extracting human contours based on 2-D photographs.

Conclusion This paper proposed a deep learning-based method for human contour extraction from complex backgrounds, use the HED network optimization model as the main framework to train and learn the human photo dataset, achieving the contour extraction and edge refinement of human photos based on complex backgrounds. Meanwhile, in order to verify the authenticity of the photo-extracted contours, 13 scale and angle parameters that can reflect the human body shape were selected, and error analysis was performed on the contour extraction values and 3-D point cloud measurements of 40 subjects. The results show that the improved HED network model can accurately extract clear and continuous human contours based on complex backgrounds with cleaner backgrounds; and there is no significant difference between the extracted values of contour-based parameters and 3-D point cloud measurements. This research proves the feasibility and accuracy of the method adopted in this research, and the research results can provide technical support for the research of non-contact 2-D measurement technology.

Key words: complex background, deep learning, human contour extraction, holistically-nested edge detection network, two-dimensional photo