一种基于空洞卷积组合的轻量级语义分割方法  

A lightweight semantic segmentation method based on combination of dilated convolution

作　　者：张晓庆 刘伟科[2] ZHANG Xiaoqing;LIU Weike(College of Computer Science and Engineering, Shandong University of Science and Technology,Qingdao,Shandong 266590,China;Network Security and Information Office, Shandong University of Science and Technology, Qingdao, Shandong 266590,China)

机构地区：[1]山东科技大学计算机科学与工程学院,山东青岛266590 [2]山东科技大学网络安全与信息化办公室,山东青岛266590

出　　处：《山东科技大学学报（自然科学版）》2021年第6期76-84,共9页Journal of Shandong University of Science and Technology(Natural Science)

摘　　要：编-解码结构的卷积神经网络是近年来出现的一类高准确率的图像语义分割方法,但是参数量大、对算力要求高的特点,束缚了其在无人驾驶、道路监控、遥感分类以及移动端物体检测等算力有限、实时性强的领域中的应用。针对以上问题,首先设计空洞卷积组合模块—NG-APC模块,通过规范空洞率,在扩大感受野的同时解决空洞卷积中的栅格问题;再利用NG-APC模块结合深度可分离卷积搭建编码-解码结构的NA-U-Net。最后利用该网络,提出一种基于空洞卷积组合的轻量级语义分割方法,在保持较高的分割准确率的同时大幅降低卷积模型的参数量和计算量。通过在公开数据集Cityscapes上进行实验,并与经典的FCN-8s、U-Net以及轻量级的ESP-Netv2、Refine Net-LW、LiteSeg进行对比,验证本方法的有效性。Convolutional neural network with encoding-decoding structure is a kind of high accuracy image semantic segmentation method that has just emerged in recent years.However,the large number of parameters and the high requirements for computing power restrict its application in fields with limited computing power and strong real-time performance,such as unmanned driving,road monitoring,remote sensing classification,mobile object detection and so on.In view of the above problems,this paper proposed a lightweight semantic segmentation method based on dilated convolution combination.Firstly,the dilated convolution combination module NG-APC module was designed by regulating the dilated rate to expand the receptive field and to solve the gridding problem in dilated convolution.Then,the NA-U-Net of encoding-decoding structure was built by using NG-APC module combined with depthwise separable convolution.With the NA-U-Net,a lightweight semantic segmentation method based on the combination of dilated convolution was put forward to maintain high segmentation accuracy and drastically reduce the amount of parameters and computation of convolution model.Extensive experiments were conducted on the Cityscapes public datasets and comparisons were made with classic FCN-8s and U-Net as well as lightweight ESP-Netv2,Refine Net-LW and Liteseg.The results verified the effectiveness of the proposed method.

关 键 词：语义分割 空洞卷积组合 感受野 栅格问题 深度可分离卷积 

分 类 号：TP30[自动化与计算机技术—计算机系统结构]