深浅交错式特征融合的人体图像分割方法  

Human image segmentation method based on depth staggered feature fusion

作　　者：帅珍彬 冯寿廷[1] SHUAI Zhenbin;FENG Shouting(School of Physics and Telecommunication Engineering,South China Normal University,Guangzhou 510006,China)

机构地区：[1]华南师范大学物理与电信工程学院,广东广州510006

出　　处：《光学技术》2020年第5期613-618,共6页Optical Technique

基　　金：国家自然科学基金重点项目(U1301251)。

摘　　要：人体图像分割作为人体行为理解和分析的基础,但要实现精准分割及实时分割是一个巨大的难题,因此提出一种深浅交错式特征融合的全卷积神经网络的方法,应用于人体图像分割。使用全卷积神经网络的卷积层提取丰富的图像特征,对不同深度的特征图由深到浅交错式地拼接并融合。最终将融合特征图送入卷积层输出预测图像,并经过全局阈值分割得到分割结果。在百度人体图像分割数据库上进行实验,其平均覆盖率可以达到89.95%,最佳分割重叠率高达99.31%;分割一幅500×500彩色图像的平均耗时为56ms,实现较好的分割性能。Human image segmentation serves as the basis for understanding and analyzing human behavior,but achieving accurate segmentation and real-time segmentation is a huge challenge.Therefore,a fully convolutional neural network method with depth staggered feature fusion is proposed for human image segmentation.The convolutional layers are utilized to extract rich features,and then feature maps of different depths are stitched and fused from deep to shallow.Finally,the fusion feature map is sent to the convolution layer to output the predicted image and the segmentation result is obtained through global threshold segmentation.Experiments on Baidu Human Image Segmentation Database show that the feature fusion structure is more efficient than the traditional method of increasing parameters,using fewer parameters to obtain more accurate segmentation results and faster segmentation speed.The average coverage rate can reach 89.95%,and the optimal segmentation result is as high as 99.31%.The average segmentation time of a 500×500 color image is 56 ms,it has better segmentation performance the segmentation performance.

关 键 词：深度学习 全卷积神经网络 特征融合 图像分割 重叠率 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]