ARSOD-YOLO: Enhancing Small Target Detection for Remote Sensing Images.

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Artificial Intelligence

Background:

• Remote sensing images are crucial for environmental monitoring, agriculture, and autonomous driving.
• Detecting small targets in these images presents significant challenges, impacting various applications.
• Existing methods require enhancement in feature extraction, fusion, and model optimization.

Purpose of the Study:

• To develop an advanced object detection model for improved small target identification in remote sensing imagery.
• To enhance feature extraction and fusion mechanisms for greater detection accuracy.
• To optimize the overall network architecture and loss functions for efficiency and performance.

Main Methods:

• Introduced the Adaptive Selective Feature Enhancement Module (AFEM) for dynamic feature weight adjustment.
• Developed the Adaptive Multi-scale Convolution Kernel Feature Fusion Module (AKSFFM) for enhanced feature fusion.
• Optimized the YOLOv8 architecture, incorporating novel component modules and loss functions, resulting in the ARSOD-YOLO model.

Main Results:

• ARSOD-YOLO demonstrated superior performance on benchmark datasets (VEDAI and AI-TOD).
• Achieved an mAP50 of 74.3% on the VEDAI dataset, a 3.1% improvement over the YOLOv8 baseline.
• Reached an mAP50 of 47.8% on the AI-TOD dataset, exceeding the baseline by 6.1%.

Conclusions:

• The proposed ARSOD-YOLO model significantly enhances small target detection capabilities in remote sensing images.
• The innovative AFEM and AKSFFM modules contribute to improved feature representation and fusion.
• ARSOD-YOLO offers a promising solution for accurate and efficient object detection in critical remote sensing applications.