A Refined and Efficient CNN Algorithm for Remote Sensing Object Detection.

Area of Science:

• Computer Vision
• Geospatial Analysis
• Artificial Intelligence

Background:

• Remote sensing object detection (RSOD) is vital for resource management, disaster assessment, and urban planning.
• Deep learning methods are effective but struggle with small, dense, or complex object arrangements in remote sensing images.

Purpose of the Study:

• To propose a refined and efficient object-detection algorithm (RE-YOLO) for improved RSOD.
• To address the challenges of detecting small, dense, and complexly arranged objects in remote sensing imagery.

Main Methods:

• Introduced a refined and efficient module (REM) within the RE_CSP block for balanced complexity and feature extraction.
• Incorporated a spatial extracted attention module (SEAM) in the backbone to enhance feature learning and semantic capture.
• Developed a three-branch path aggregation network (TBPAN) for comprehensive fusion of multi-scale and multi-channel information.

Main Results:

• The RE-YOLO algorithm demonstrated superior performance compared to state-of-the-art methods on DOTA-v1.0 and SCERL datasets.
• Achieved significant improvements in generalization ability for remote sensing object detection tasks.
• Effectively handled challenges related to object size, density, and arrangement.

Conclusions:

• RE-YOLO offers a robust and efficient solution for remote sensing object detection.
• The proposed modules (REM, SEAM, TBPAN) contribute to enhanced feature representation and contextual understanding.
• The algorithm shows strong potential for practical applications in resource utilization and urban planning.