SAMF-YOLO: A self-supervised, high-precision approach for defect detection in complex industrial environments.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Balancing computational efficiency and feature expressiveness is crucial for complex object detection models.
• Existing models face challenges in detecting small objects and handling scale variations.

Purpose of the Study:

• To introduce SAMF-YOLO, a novel object detection model designed for enhanced accuracy and efficiency.
• To improve the detection of small defects and address challenges in bounding box regression.

Main Methods:

• Integration of SONet, Bi-temporal Feature Aggregation Module (BFAM), and FASFF-Head within a UniRepLKNet backbone enhanced by Star Operation.
• Utilizing Focaler-IoU loss for improved bounding box regression and self-supervised contrastive learning for feature representation.
• Adaptive multi-scale feature fusion with minimal computational overhead.

Main Results:

• SAMF-YOLO achieved a 6.38% improvement in mAP@0.5 compared to YOLOv11s.
• Demonstrated a significant reduction in computational cost while maintaining high accuracy.
• Showcased enhanced robustness and superior detection of small defects.

Conclusions:

• SAMF-YOLO offers a superior balance of accuracy, efficiency, and robustness in object detection.
• The proposed model effectively addresses limitations in existing object detection architectures.
• The integration of novel modules and loss functions contributes to state-of-the-art performance.