基于改进的IIE-SegNet的快速图像语义分割方法  

Fast image semantic segmentation method based on improved IIE-SegNet

作　　者：李庆[1,2] 王宏健 李本银 肖瑶 迟志康[1] LI Qing;WANG Hongjian;LI Benyin;XIAO Yao;CHI Zhikang(College of Intelligent Systems Science and Engineering,Harbin Engineering University,Harbin 150001,China;College of Intelligent Science and Engineering,Yantai Nanshan University,Yantai 264000,China)

机构地区：[1]哈尔滨工程大学智能科学与工程学院,黑龙江哈尔滨150001 [2]烟台南山学院智能科学与工程学院,山东烟台264000

出　　处：《哈尔滨工程大学学报》2024年第2期314-323,共10页Journal of Harbin Engineering University

基　　金：GF科技创新特区项目(21-163-05-ZT-002-005-03);水下机器人重点实验室基金项目(JCKYS2022SXJQR-09);哈尔滨工程大学“高水平科研引导专项”(3072022QBZ0403)。

摘　　要：针对IIE-SegNet计算复杂度高、计算量大等问题,本文提出一种基于IIE-SegNet的改进方法。编码结构中引入经ImageNet训练过的VGG16和多尺度空洞卷积空间金字塔池化来获得丰富的编码信息;解码结构中,设计全局加平均模块来解决IIE-SegNet计算量大的问题;研究Focal损失函数来解决正、负采样不平衡的问题。实验结果表明:与IIE-SegNet相比,本方法在PASCAL VOC 2012数据集上的语义分割速度更快,平均每次迭代快0.6 s左右,测试单张图像的时间平均减少了0.94 s;分割精度更高,MIoU提升了2.1%。在扩展的PASCAL VOC 2012(Exp-PASCAL VOC 2012)数据集上的语义分割速度更快,平均每次迭代快0.4 s左右,测试单张图像的时间平均减少了0.92 s;分割精度更高,MPA和MIoU分别提升了2.6%和2.8%,特别是对于小尺度目标分割边界更清晰,性能得到了很大的提升。To address the high computational complexity and large computational load of IIE-SegNet,this paper proposes an improved method based on IIE-SegNet.VGG16 trained in ImageNet and multiscale atrous spatial pyra-mid pooling(MASPP)module are introduced into the encoding module to obtain abundant coding information.In the decoding structure,the global add average(GAA)module is designed to solve the problem of heavy computa-tion of IIE-SegNet.Focal loss function is analyzed to solve the problem of positive and negative sampling imbal-ance.The experimental result shows that compared with IIE-SegNet,on the PASCAL VOC 2012 dataset,our net-work achieves faster segmentation,with an average of 0.6 s faster per iteration,a reduction in the average time to test a single image by 0.94 s,and an increase in MIoU 2.1%.On the expanded PASCAL VOC 2012(Exp-PAS-CAL VOC 2012)data set,the semantic segmentation speed is faster,an average of 0.4 s faster per iteration,and the average time to test a single image is reduced by 0.92 s;furthermore,our network exhibits higher accuracy and improvements in the MPA and MIoU by 2.6%and 2.8%,respectively.In particular,for small-scale target seg-mentation,the boundary is clearer,and the performance is greatly improved.

关 键 词：语义分割 深度学习 多尺度空洞卷积空间金字塔池化 图像信息熵 全局加平均 VGG16 IIE-SegNet 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]