Enhancing Semi-Supervised Semantic Segmentation of Remote Sensing Images via Feature Perturbation-Based Consistency...

Area of Science:

• Remote Sensing Technology
• Computer Vision
• Machine Learning

Background:

• Semantic segmentation of remote sensing images is crucial but hindered by the time-consuming nature of manual data annotation.
• Existing semi-supervised methods struggle with maintaining prediction consistency due to complex foreground categories and feature spaces in remote sensing data.
• Optimization within complex feature spaces is challenging, leading to confusion between different categories.

Purpose of the Study:

• To propose a novel semi-supervised semantic segmentation framework to enhance model consistency and optimize feature-based class categorization for remote sensing images.
• To address the limitations of conventional Mean Teacher by incorporating feature-level perturbations.
• To improve the accuracy of semantic segmentation in complex remote sensing image datasets.

Main Methods:

• A novel semi-supervised semantic segmentation framework based on the Mean Teacher (MT) model.
• Introduction of feature-level perturbations, in addition to image-level perturbations, to the Mean Teacher framework.
• Application of contrastive learning for feature-level learning to maintain consistency after feature perturbation.
• Utilization of an entropy threshold to assist contrastive learning in precisely selecting feature key-values for complex remote sensing image features.

Main Results:

• The proposed framework demonstrates superior performance compared to existing methods on benchmark datasets.
• Enhanced prediction consistency was achieved through feature-level perturbations and contrastive learning.
• Improved accuracy in semantic segmentation of complex remote sensing images was observed.
• The entropy threshold effectively assisted contrastive learning in refining feature selection.

Conclusions:

• The developed semi-supervised semantic segmentation framework effectively addresses the challenges of complex features and prediction consistency in remote sensing images.
• Feature-level perturbations combined with contrastive learning and entropy thresholding offer a robust approach for improving segmentation accuracy.
• The methodology shows significant potential for advancing automated analysis of remote sensing data.