Omni-dimensional dynamic convolution with coordinate attention detection scheme.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Small object detection remains a challenge in computer vision due to low resolution and fine-grained details.
• Existing models like YOLOv8n require architectural modifications to effectively capture features of small targets.

Purpose of the Study:

• To improve the small target detection capabilities of the YOLOv8n model.
• To enhance spatial feature representation and multi-scale feature fusion for better small object recognition.

Main Methods:

• Incorporated coordinate attention (CA) into the C2f module to refine spatial focus.
• Replaced the Path Aggregation Network with Bidirectional Feature Pyramid Network (BiFPN) for superior multi-scale feature fusion.
• Introduced an additional smaller detection head with Omni-dimensional Dynamic Convolution (ODConv) for enhanced perception of very small objects.

Main Results:

• Achieved significant improvements in small object detection metrics, including average precision mean (mAP), precision, and recall.
• Demonstrated a 3.2% increase in mAP@50 and 4.4% in mAP@75 for small targets compared to the original YOLOv8n.
• Showcased enhanced ability to capture fine-grained features and address challenges like low contrast and scale variations.

Conclusions:

• The proposed enhancements effectively improve YOLOv8n's performance on small object detection tasks.
• The integration of CA, BiFPN, and ODConv offers a robust solution for recognizing small objects in complex scenarios.
• This work contributes to advancing the accuracy and reliability of object detection systems for applications requiring the identification of diminutive targets.