NRT-YOLO: Improved YOLOv5 Based on Nested Residual Transformer for Tiny Remote Sensing Object Detection.

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Artificial Intelligence

Background:

• Object detection in high-resolution remote sensing imagery faces challenges with tiny objects.
• Existing coarse-grained cropping methods are computationally intensive and complex.
• There's a need for efficient and accurate detection models for small objects in remote sensing.

Purpose of the Study:

• To develop an improved object detection architecture, NRT-YOLO, for enhanced performance in remote sensing.
• To address the limitations of existing methods in detecting tiny objects and high-resolution imagery.
• To reduce computational complexity while improving detection accuracy.

Main Methods:

• An improved YOLO architecture, NRT-YOLO, was proposed.
• Key improvements include an extra prediction head, feature fusion layers, a novel nested residual Transformer module (C3NRT), and a nested residual attention module (C3NRA).
• Multi-scale testing was employed, and ablation studies validated the effectiveness of C3NRT.

Main Results:

• NRT-YOLO achieved 56.9% mAP0.5 on the DOTA dataset with only 38.1 M parameters, outperforming YOLOv5l by 4.5%.
• The C3NRT module demonstrated the largest accuracy improvement (2.7% mAP0.5).
• The model showed excellent capability in detecting small objects across different classifications.

Conclusions:

• NRT-YOLO offers significant improvements in accuracy and parameter reduction for tiny object detection in remote sensing.
• The proposed C3NRT module effectively boosts accuracy and reduces network complexity.
• NRT-YOLO is a suitable and efficient solution for remote sensing object detection tasks involving small objects.