Weakly Supervised Semantic Segmentation of Remote Sensing Images Based on SD-CAM with Improved Mumford-Shah Loss

Weakly supervised semantic segmentation (WSSS) based on image-level annotation has received wide attention in the field of multi-class remote sensing images(RSIs). However, there is a lack of background information and there are a wide variety of targets in RSIs, which makes WSSS of RSIs suffer from incomplete information on small targets in class activation maps (CAMs). The WSSS of RSIs based on SD-CAM with improved mumford-shah loss (IMS-Loss) is proposed to solve the above challenge. The multi-stage WSSS method is adopted to improve the segmentation accuracy in novel method. Firstly, the self-attention mechanism (SAM) module is introduced into different stages of classification network to dig into the detail information of the RSI in the generation CAMs stage. Secondly, the feature maps are processed by the SAM module and are fed to the multi-dilated rate convolution (MDRC) module to obtain high quality CAMs, which enlarges the area of the localization target of the CAMs to improve the pseudo-labeling accuracy. Then, the improved Mumford-Shah Loss (IMS-loss) function is introduced to provide the segmentation network with supervised information of the original images to produce high quality segmentation results in the semantic segmentation network. Finally, simulation results show that the improved method is capable of obtaining high-precision segmentation results using image-level annotation information. The mIOU of the novel method on Vaihingen, Potsdam and WHDLD datasets reach 85.58%, 86.39% and 70.03% of the semantic segmentation methods for fully supervised semantic segmentation (FSSS) of RSIs.