HSF-YOLO: A Multi-Scale and Gradient-Aware Network for Small Object Detection in Remote Sensing Images.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Small object detection (SOD) in remote sensing images (RSIs) faces challenges like scale variation, occlusion, and complex backgrounds.
• Existing methods often result in high miss and false detection rates for small objects.

Purpose of the Study:

• To propose a novel detection framework, HSF-YOLO, for enhanced small object detection in RSIs.
• To improve feature encoding, attention interaction, and localization precision within the YOLOv8 backbone.

Main Methods:

• Introduced Hybrid Atrous Enhanced Convolution (HAEC) to capture multi-scale semantic and local information.
• Implemented a Spatial-Interactive-Shuffle attention module (C2f_SIS) for enhanced feature interaction and noise suppression.
• Developed Focal Gradient Refinement Loss (FGR-Loss) for improved regression accuracy and training robustness.

Main Results:

• HSF-YOLO demonstrated significant improvements over the baseline YOLOv8 on three public datasets (VisDrone2019, DIOR, NWPU VHR-10).
• Achieved notable gains in mAP@0.5 (up to 5.7%) and mAP@0.5:0.95 (up to 4.0%) on the VisDrone2019 dataset.
• Showcased consistent performance improvements across different RSI datasets, confirming effectiveness.

Conclusions:

• HSF-YOLO provides a unified and effective solution for small object detection in complex RSI scenarios.
• The framework offers a favorable balance between detection accuracy and computational efficiency.
• The proposed modules contribute to overcoming the limitations of current SOD techniques in RSIs.