Region-Enhancing Network for Semantic Segmentation of Remote-Sensing Imagery

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Semantic segmentation of high-resolution remote-sensing imagery (HRRSI) is crucial for machine vision applications.
• Current deep learning methods often overemphasize pixel-level details, leading to overlearning and misclassification due to annotation errors and complex features.

Purpose of the Study:

• To introduce a novel semantic segmentation network, the region-enhancing network (RE-Net), designed to overcome the limitations of pixel-based approaches.
• To enhance the understanding of regional integrity and context in HRRSI for improved segmentation accuracy.

Main Methods:

• RE-Net integrates regional information into a base network, utilizing a regional context learning procedure (RCLP) to capture region-level context.
• A region correcting procedure (RCP) recalibrates pixel features using aggregated regional features.
• An intra-network multi-scale attention module adaptively selects features based on region size.

Main Results:

• Extensive experiments on four public datasets demonstrate RE-Net's superior performance compared to existing state-of-the-art methods.
• The proposed approach effectively reduces misclassification by leveraging regional information and context.

Conclusions:

• RE-Net offers a more robust and accurate solution for semantic segmentation of HRRSI by prioritizing regional understanding.
• The network's ability to learn regional context and recalibrate pixel features provides significant advantages over pixel-centric methods.