基于非下采样轮廓波变换的遥感地物分割算法  

A Remote Sensing Ground Object Segmentation Algorithm Based on Non-subsampled Contourlet Transform

作　　者：闵锋[1] 彭伟明 况永刚 毛一新 郝琳琳 MIN Feng;PENG Weiming;KUANG Yonggang;MAO Yixin;HAO Linlin(Hubei Province Key Laboratory of Intelligent Robot,Wuhan Institute of Technology,Wuhan 430000,China)

机构地区：[1]武汉工程大学智能机器人湖北省重点实验室,武汉430000

出　　处：《电光与控制》2023年第11期49-55,共7页Electronics Optics & Control

基　　金：国家自然科学基金(62171328)。

摘　　要：针对遥感地物图像具有背景复杂且种类众多的特点,利用传统算法进行分割会导致边缘模糊、信息丢失及分割精度低的问题,提出了一种基于改进DeepLabV3+网络的语义分割算法。首先,在主干网络中引入改进后的特征提取网络CHRNet;其次,使用非下采样轮廓波变换(NSCT)算法重构空洞空间金字塔池化(ASPP)模块中的全局池化操作;最后,在模型编码和解码阶段添加无参数的注意力机制SimAM,加强模块间的特征传递,提高特征利用率。实验表明,在PASCAL VOC2012和WHDLD数据集上,改进算法的平均交并比(MIoU)分别达到了81.56%和64.2%,较原有算法分别提升了约4.61和2.8个百分点,改进算法在保证分割速率的同时,提升了分割精度。Remote sensing ground object images have the characteristics of complex background and numerous varieties,and traditional segmentation algorithms will lead to edge blur,information loss and low segmentation accuracy.To solve the problems,a semantic segmentation algorithm based on the improved DeepLabV3+network is proposed.Firstly,the improved feature extraction network CHRNet is introduced into the backbone network.Secondly,the Non-Subsampled Contourlet Transform(NSCT)algorithm is used to reconstruct the global pooling operation in the Atrous Spatial Pyramid Pooling(ASPP)module.Finally,the parameter-free attention mechanism SimAM is added in model encoding and decoding stages to enhance feature transfer among modules and improve feature utilization ratio.The experimental results show that Mean Intersection over Union(MIoU)of the improved algorithm is 81.56% on PASCAL VOC2012 data set and 64.2% on WHDLD data set,which are about 4.61 percentage points and 2.8 percentage points higher than those of the original algorithm.The improved algorithm can enhance segmentation accuracy while ensuring segmentation speed.

关 键 词：遥感图像 DeepLabV3+ 非下采样轮廓波变换 空洞空间金字塔池化 注意力机制 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]