一种轻量化多尺度遥感图像分割方法

A Lightweight Multi-scale Remote Sensing Image Segmentation Method

遥感图像的语义分割在城市规划和发展中发挥着至关重要的作用。如何对高复杂度、多类别的遥感影像进行自动、快速、有效的语义分割已成为研究的关键。现有的基于深度学习的分割方法存在模型复杂、计算成本较高等问题。提出一种端到端的轻量级多尺度特征提取分割网络(Multi-Scale Feature Extraction and Segmentation Network MSNET),旨在解决在高准确性情况下降低计算成本的问题。主干是基于轻量级网络MobileNetV2的编码网络和基于MSConv的解码网络构成的整个主干,其中MSConv是一种新的多尺度卷积模块。还提出了一种特征融合注意力模块(Feature Fusion Attention Module, MSAM)来有效地整合通道和空间维度上注意机制的全局信息。引入更加轻量化的局部重要性池化(Local Importance Pooling, LIP)代替普通池化操作以及添加了空洞空间卷积池化金字塔(Atrous Spatial Pyramid Pooling, ASPP)模块进一步提取丰富的特征。在公开数据集WHDLD上进行对比评估,F1达到83.12%,推理时间仅为0.007 4 s。

Semantic segmentation of remote sensing images plays an important role in urban planning and development.How to perform automatic,fast and effective semantic segmentation for highly complex and multi-class remote sensing images has become the key of research.However,the existing segmentation methods based on deep learning have the problems of complex model and high computational cost.An end-to-end lightweight Multi-Scale Feature Extraction and Segmentation Network(MSNET)is proposed to reduce the computational cost in the case of high accuracy.Firstly,the backbone is composed of a coding network based on lightweight network MobileNetV2 and a decoding network based on MSConv.MSConv is a new multi-scale convolutional module.In addition,a Feature Fusion Attention Module(MSAM)is proposed to effectively integrate the global information of attention mechanisms in channel and spatial dimensions.Secondly,a more lightweight Local Importance Pooling(LIP)is introduced to replace the common pooling operation,and the Atrous Spatial Pyramid Pooling(ASPP)module is added to further extract rich features.Finally,the comparison evaluation on the public dataset WHDLD showed that F1 reaches 83.12%and the inference time is only 0.0074 s.