DBA-YOLO: A Dense Target Detection Model Based on Lightweight Neural Networks

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Industrial deep learning models struggle with computational complexity and accuracy for dense, small targets.
• Existing models often fail in real-world scenarios due to missed/false detections from complex backgrounds and inter-class similarity.

Purpose of the Study:

• To develop DBA-YOLO, a lightweight model based on YOLOv10, that enhances inference efficiency and detection accuracy for industrial applications.
• To address challenges of high computational complexity, small target detection, and complex background interference in dense target detection.

Main Methods:

• Proposed DBA-YOLO, a lightweight model integrating a C2f PA module for feature extraction, a parameter-refined BIMAFPN neck for small target detection, and a DyDHead module for spatial feature weighting.
• Utilized model compression and algorithm optimization techniques to reduce computational complexity.
• Constructed a real-world dataset of cigarette package images for validation.

Main Results:

• DBA-YOLO achieved 91.3% detection accuracy on SKU-110K and the custom dataset, outperforming the baseline by 1.4%.
• Demonstrated 2-3% improvements in mean Average Precision (mAP) and mAP75.
• Reduced model parameters by 3.6%, indicating enhanced efficiency.

Conclusions:

• DBA-YOLO effectively balances high detection performance and computational efficiency, making it suitable for resource-constrained industrial devices.
• The proposed model offers a viable solution for accurate and efficient dense target detection in challenging industrial environments.