A new lightweight deep neural network for surface scratch detection.

Area of Science:

• Materials Science and Engineering
• Artificial Intelligence and Machine Learning

Background:

• Surface scratches on components in contact sliding, such as in metal forming, can significantly impact performance and lifespan.
• Accurate and efficient detection of these scratches is crucial for quality control and predictive maintenance.

Purpose of the Study:

• To develop a lightweight convolutional neural network (CNN) named WearNet for automated surface scratch detection.
• To evaluate WearNet's performance against existing methods in terms of accuracy, model size, and detection speed.

Main Methods:

• Training WearNet using a large dataset of surface scratches generated from cylinder-on-flat sliding tests.
• Investigating WearNet's network response and decision-making mechanisms.
• Evaluating WearNet on a public image database and an embedded system for sheet metal forming applications.

Main Results:

• WearNet achieved a classification accuracy of 94.16%.
• The developed network demonstrated a significantly smaller model size and faster detection speed compared to existing CNNs.
• WearNet outperformed other state-of-the-art networks on a public image database and showed practical advantages in an embedded system.

Conclusions:

• WearNet is an effective and efficient solution for automatic surface scratch detection in critical industrial applications like metal forming.
• The lightweight design of WearNet makes it suitable for deployment in resource-constrained embedded systems.
• The study highlights the potential of specialized CNNs for improving quality control and reducing material defects.