An efficient and lightweight algorithm for detecting surface defects of steel based on SCCI-YOLO

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Steel surface defects critically impact industrial product quality.
• Existing deep learning methods struggle with diverse, complex, and multi-scale defect detection, leading to inaccuracies.
• Limitations include poor feature extraction, fusion, and multi-scale recognition.

Purpose of the Study:

• To develop an advanced steel surface defect detection algorithm overcoming current deep learning limitations.
• To improve accuracy, reduce false detection rates, and enhance recognition of multi-scale defects.
• To propose a novel model, SCCI-YOLO, based on an improved YOLOv8n architecture.

Main Methods:

• Integrated the SPD-Conv module into the backbone for adaptive convolutional kernel focus, enhancing small object detection.
• Developed the C2f_EMA module for improved feature extraction and fusion.
• Incorporated a lightweight cross-scale feature fusion module (CCFM) in the Neck network for multi-scale adaptability.
• Utilized the Inner-IoU loss function to improve model convergence and regression accuracy.

Main Results:

• SCCI-YOLO achieved a mean Average Precision (mAP) of 78.6% on the NEU-DET dataset.
• Demonstrated a 2.2% and 5.9% improvement in detection accuracy over YOLOv8n and YOLOv7, respectively.
• Reduced model parameters by 43.9% compared to the original YOLOv8n model.

Conclusions:

• The SCCI-YOLO algorithm shows excellent overall performance in steel surface defect detection.
• The proposed improvements effectively address feature extraction, fusion, and multi-scale recognition challenges.
• The model offers a more accurate and efficient solution for industrial steel surface quality control.