Centralised visual processing center for remote sensing target detection.

Area of Science:

• Remote Sensing and Computer Vision
• Artificial Intelligence for Geospatial Analysis

Background:

• Accurate target detection in satellite imagery is challenged by scale variations, dense distributions, complex backgrounds, and occlusion.
• Existing methods struggle with high target feature similarity and severe occlusion in overhead remote sensing images.

Purpose of the Study:

• To develop a novel network for efficient and accurate target detection in remote sensing images.
• To address challenges of scale, density, feature similarity, and occlusion in satellite imagery analysis.

Main Methods:

• Proposed a Centralised Visual Processing Center (CVPC) integrating Transformer encoder and CNN for broad and detailed feature capture.
• Introduced a Pixel-level Learning Center (PLC) module to enhance pixel-level correlations and feature depiction.
• Developed a centralized feature cross-layer fusion pyramid and a Context Enhanced Adaptive Sparse Convolutional Network (CEASC).

Main Results:

• The proposed 3CNet achieved advanced detection accuracy on challenging datasets: DOTA-v1.0 (78.62% mAP), DIOR (79.12% mAP), and RSOD (95.50% mAP).
• The CEASC network demonstrated a balance between high detection accuracy and efficient processing speed.

Conclusions:

• The developed 3CNet, incorporating CVPC, PLC, and CEASC, significantly enhances target detection performance in remote sensing.
• The approach effectively tackles issues of feature similarity and occlusion, paving the way for improved satellite image analysis.