CRABR-Net: A Contextual Relational Attention-Based Recognition Network for Remote Sensing Scene Objective.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Remote sensing scene objective recognition (RSSOR) is crucial for military and civilian applications.
• Convolutional Neural Networks (CNNs) have advanced RSSOR, but current methods often ignore inter-layer feature relationships, leading to suboptimal accuracy.
• Existing CNN-based RSSOR techniques frequently use only the last layer's features or simple fusion methods, causing feature redundancy and loss.

Purpose of the Study:

• To develop an advanced CNN-based network for high-resolution remote sensing scene objective recognition.
• To address limitations in feature fusion and inter-layer relationship utilization in existing RSSOR methods.
• To enhance the feature learning capability for more accurate and robust objective recognition in remote sensing images.

Main Methods:

• Introduced the Contextual, Relational Attention-based Recognition Network (CRABR-Net).
• Employed a parameter-free attention module (SimAM) to focus on salient feature content from different CNN layers.
• Implemented complementary and enhanced relationship feature map calculations to effectively fuse adjacent feature maps and improve feature learning.
• Utilized concatenated feature maps from multiple layers for the final RSSOR.

Main Results:

• CRABR-Net effectively exploits relationships between different CNN layers to boost recognition performance.
• The proposed network achieved superior results compared to several state-of-the-art algorithms.
• Achieved high average accuracies: 96.46% on AID, 99.20% on UC-Merced, and 95.43% on RSSCN7 with standard training ratios.

Conclusions:

• CRABR-Net demonstrates the significant benefit of leveraging inter-layer feature relationships in CNNs for RSSOR.
• The novel feature fusion and attention mechanisms contribute to improved accuracy and efficiency in remote sensing image analysis.
• The network offers a promising approach for enhancing the intelligence of objective recognition in high-resolution remote sensing scenes.