Enhanced YOLOv8 for industrial polymer films: a semi-supervised framework for micron-scale defect detection

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Polymer film production via extrusion can introduce micro-defects impacting product quality.
• Traditional machine learning methods for defect detection lack accuracy and adaptability.

Purpose of the Study:

• To propose an improved YOLOv8 model for accurate micro-defect detection on polymer films.
• To enhance detection of small targets and improve generalization capabilities.

Main Methods:

• Implemented an improved YOLOv8 model incorporating the CBAM attention mechanism.
• Utilized semi-supervised learning with the Mean Teacher method and limited labeled data.
• Employed an improved loss function based on normalized Wasserstein distance (NWD) and a multi-threshold mask segmentation algorithm.

Main Results:

• Achieved an 8.26% increase in mAP@0.5 compared to the baseline YOLOv8 model.
• Demonstrated high precision for small target detection, exceeding 95.0% detection rates across various image sizes.
• Showcased superior performance over traditional methods in varied lighting and contamination conditions.

Conclusions:

• The enhanced YOLOv8 algorithm meets stringent industrial requirements for high-precision small-target defect detection in polymer films.
• Future research will focus on further enhancing model accuracy and robustness for defect detection.