Semantic Segmentation of Remote Sensing Images Using the Channel Domain
Attention Mechanism Deeplabv3+ Algorithm
Author information+
{{custom_zuoZheDiZhi}}
{{custom_authorNodes}}
{{custom_bio.content}}
{{custom_bio.content}}
{{custom_authorNodes}}
Collapse
History+
Published
2023-02-20
Issue Date
2025-06-08
Abstract
In order to extract typical ground objects of high-resolution remote sensing images, and to solve the
problems of fuzzy edge segmentation in Deeplabv3+ remote sensing image segmentation, such as holes and
missing classification, a method of adding channel attention mechanism module to Deeplabv3+ is proposed
based on deep learning to enhance segmentation results. First, the high-level feature map obtained through the
deep convolutional network is input to the channel attention mechanism, and pixel-level feature enhancement
between channels is performed. Then the multi-scale input image is obtained through the spatial pyramid pool,
and the category imbalance corrected to extract the completed segmentation information of the image realizes
the optimization of segmentation boundary information. Finally, by collecting remote sensing images of
Guangzhou City Gaofen No. 2 for remote sensing data processing, labeling, and after enhancement,
experiments are carried out and compared with classic semantic segmentation networks U-Net, SegNet, and
PSPNet. The results show that the evaluation index MIOU of this method reaches by 96.19%, MPA reaches by
97.85%.
XU Chang-you, FAN Shao-sheng, ZHU Hang.
Semantic Segmentation of Remote Sensing Images Using the Channel Domain
Attention Mechanism Deeplabv3+ Algorithm[J]. Control Engineering of China, 2023, 30(2): 368-375