Context Enhancing Representation for Semantic Segmentation in Remote Sensing Images.

Area of Science:

• Remote Sensing
• Computer Vision
• Image Analysis

Background:

• Semantic segmentation is crucial for remote sensing image interpretation.
• Existing methods struggle with multiscale objects, leading to fragmented large objects, missed small objects, and blurred boundaries due to insufficient context modeling.

Purpose of the Study:

• To propose a Contextual Representation Enhancement Network (CRENet) for improved semantic segmentation in remote sensing images.
• To enhance the modeling of global context (GC) and local context (LC) in high-level features.

Main Methods:

• The CRENet utilizes a Local Feature Alignment Enhancement Module (LFAEM) for aligning and enhancing local context using cascaded deformable convolution.
• A Superpixel Affinity Loss (SAL) is employed to improve global context capture by supervising semantic information from superpixels.
• The method is designed as a plug-and-play module compatible with FCN-based networks.

Main Results:

• Experiments on two remote sensing image datasets demonstrate the effectiveness of CRENet.
• The proposed network achieves competitive performance in both qualitative and quantitative evaluations.
• CRENet successfully addresses issues of object fragmentation, small object detection, and boundary blurring.

Conclusions:

• The CRENet effectively enhances global and local context modeling for semantic segmentation in remote sensing.
• The proposed LFAEM and SAL components contribute to improved accuracy and robustness.
• CRENet offers a versatile solution for enhancing existing FCN-based semantic segmentation models.