ECM-YOLO: a real-time detection method of steel surface defects based on multiscale convolution.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Steel surface defects negatively impact product quality, performance, and reliability.
• Accurate and high-speed detection of diverse steel defects is critical for industrial applications.

Purpose of the Study:

• To develop a novel, high-precision, and high-speed steel surface defect detection algorithm.
• To enhance existing deep learning models for improved defect identification and classification.

Main Methods:

• Proposed ECM-YOLO detection network based on YOLOv8n.
• Introduced C2f enhanced multiscale convolution processing (C2f_EMSCP) module for improved feature capture.
• Integrated channel prior convolutional attention (CPCA) mechanism for efficient information transmission.
• Developed a multiscale simple and efficient anchor matching head (MultiSEAMHead) to address overlapping defects.

Main Results:

• ECM-YOLO achieved mAPs of 78.9% on NEU-DET and 68.2% on GC 10-DET datasets.
• Demonstrated superior performance over YOLOv8n, with improvements of 2.5% and 4.4% respectively.
• Exhibited advantages in model parameters, computational efficiency, and inference speed compared to other models.

Conclusions:

• ECM-YOLO offers a robust solution for real-time steel surface defect detection.
• The proposed modules enhance feature extraction and defect localization accuracy.
• ECM-YOLO is applicable for improving quality control in industrial settings.