MDI-YOLO a lightweight transformer-CNN-based multidimensional feature fusion model for small object detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Object detection in aerial and remote sensing images faces challenges like small target size, background interference, and occlusion.
• Existing models often struggle with accuracy due to these complex environmental factors.

Purpose of the Study:

• To develop a lightweight and accurate small object detection model for drone aerial and remote sensing images.
• To enhance feature extraction and object localization capabilities in challenging visual conditions.

Main Methods:

• Proposes MDI-YOLO, a model integrating Transformer and Convolutional Neural Network (CNN) features.
• Introduces C2f-MCC structure for improved global feature extraction and Directional Fusion Attention (DFA) for enhanced feature representation.
• Implements Inner-Shape-IoU loss function for precise bounding box evaluation.

Main Results:

• Achieved a 4% increase in mAP@0.5 and 2.5% in mAP@0.5:0.95 on the VisDrone2019 dataset.
• Improved mAP@0.5 by 3.3% and mAP@0.5:0.95 by 2.8% on the DOTAv1.0 dataset.
• Maintained lightweight characteristics with minimal parameter increase.

Conclusions:

• MDI-YOLO effectively enhances small object detection accuracy in drone imagery.
• The model demonstrates robustness and generalization, suitable for aerial and remote sensing applications.
• The proposed methods offer a significant advancement in lightweight, high-performance object detection.