OD-YOLO: Robust Small Object Detection Model in Remote Sensing Image with a Novel Multi-Scale Feature Fusion

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Remote sensing technology is vital for hydrology, agriculture, and geography, but object detection accuracy is hindered by expansive scenes and small, dense targets.
• Existing models struggle with feature extraction for geometric shapes and blurred objects, limiting performance in remote sensing imagery.
• Developing advanced models is crucial to overcome these limitations and improve object identification capabilities.

Purpose of the Study:

• To design an improved object detection model, OD-YOLO, specifically for enhancing the identification of small objects in remote sensing imagery.
• To address the limitations of traditional convolutional neural networks in recognizing geometric shapes and blurred objects.
• To boost the performance of the YOLOv8n model through multi-scale feature fusion and novel module integration.

Main Methods:

• Introduced the Detection Refinement Module (DRmodule) with Deformable Convolutional Networks and Hybrid Attention Transformer for enhanced feature extraction.
• Integrated a Dynamic Head into the YOLO Feature Pyramid Network to improve the fusion of multi-scale features.
• Developed the OIoU loss function to precisely differentiate detection and true boxes, optimizing small object detection.

Main Results:

• OD-YOLO achieved superior performance on the VisDrone dataset, outperforming existing models by at least 5.2% in mAP50 and 4.4% in mAP75.
• On the Foggy Cityscapes dataset, OD-YOLO demonstrated a 6.5% improvement in mAP, showcasing effectiveness in adverse weather conditions.
• The model exhibited outstanding results in remote sensing image analysis and object detection tasks, including those in adverse weather.

Conclusions:

• The OD-YOLO approach effectively enhances object detection accuracy in remote sensing imagery, especially for small and challenging targets.
• The integration of DRmodule, Dynamic Head, and OIoU loss function provides significant improvements over baseline models.
• This research advances remote sensing image analysis and offers practical technical support for future remote sensing applications.