Enhancing object detection in remote sensing images with improved YOLOv8 model.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Object detection in remote sensing images faces challenges due to complex backgrounds, diverse appearances, and varying object scales.
• Densely distributed targets and scale variations complicate accurate detection in aerial imagery.
• Existing models struggle with the intricate nature of remote sensing data.

Purpose of the Study:

• To enhance the YOLOv8 model for improved object detection in remote sensing applications.
• To address challenges like complex backgrounds, scale variation, and dense object distribution.
• To boost detection accuracy and efficiency in remote sensing image analysis.

Main Methods:

• Integration of dynamic convolution (DyConv) in the C2F module to adaptively adjust filters for scale and appearance variations.
• Implementation of Dual level Routing Attention (BRA) to refine high-level features, suppress background noise, and enhance feature correlation.
• Adoption of Asymptotic Feature Pyramid Network (AFPN) for superior multi-scale feature fusion, combining low-level details with high-level semantics.

Main Results:

• The enhanced YOLOv8 model achieved a mAP50-95 score of 65.4% on the Remote Sensing Object Detection (RSOD) dataset, a 3.3% improvement over the original model.
• Demonstrated significant performance gains compared to mainstream single-stage, two-stage, and DETR object detection models.
• Maintained computational efficiency while improving detection accuracy.

Conclusions:

• The proposed enhancements effectively address the complexities of object detection in remote sensing imagery.
• The integration of DyConv, BRA, and AFPN offers a robust solution for accurate and efficient remote sensing object detection.
• The improved YOLOv8 model represents a significant advancement in the field of remote sensing image analysis.