Deep Bilateral Filtering Network for Point-Supervised Semantic Segmentation in Remote Sensing Images.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Deep neural networks excel at semantic segmentation but require extensive pixel-level annotations, which are labor-intensive for remote sensing data.
• Existing point-based weakly supervised methods struggle with incomplete object structures and lost boundary information due to sparse annotations.

Purpose of the Study:

• To propose a point-based weakly supervised learning framework, the Deep Bilateral Filtering Network (DBFNet), for semantic segmentation of remote sensing images.
• To address the challenges of sparse point annotations, including incomplete object prediction and loss of boundary details.

Main Methods:

• Developed a novel framework, DBFNet, incorporating deep bilateral filtering (DBF) into learned representations.
• DBF filters deep features to enhance feature consistency in smooth regions and preserve boundary information by distinguishing features across edges.
• Utilized sparse point annotations as input for training.

Main Results:

• DBFNet achieved highly competitive performance on the ISPRS 2-D semantic labeling Potsdam and Vaihingen datasets.
• The proposed method demonstrated effectiveness in overcoming the limitations of point-based annotations for semantic segmentation.
• DBFNet achieved performance comparable to state-of-the-art fully-supervised methods.

Conclusions:

• DBFNet offers an effective solution for semantic segmentation of remote sensing images using weakly supervised point annotations.
• The integration of deep bilateral filtering significantly improves prediction accuracy and boundary preservation.
• The framework provides a viable alternative to costly pixel-level annotation for large-scale remote sensing applications.