Multi-task Learning Based on Deep Convolutional Neural Networks for Surface Defect Detection of Metal Gaskets

Chao-Ching Ho, Chun-Han Liu, Ming-Fu Chen, Ming-Chieh Kao

Abstract

<p>This paper proposes a multi-task learning system for industrial defect detection, focusing on surface scratches on titanium gaskets as the detection target. Our model is built using a deep convolutional neural network (CNN). As defects are infrequent in actual production lines, we adopt the Faster-RCNN object detection network architecture and integrate it with a multi-task learning system. A classification task model is introduced into the backbone network to filter out a significant number of defect-free images. This step reduces the computation and data transmission time of the subsequent target detection task, accelerating the detection process. Experimental results show that our proposed method reduces inference time by 37.6% and 42.46% in actual production lines with defect rates of 12% and 5%, respectively, while maintaining 96% of the original model's performance.</p>

Keywords

Deep Convolutional Neural Networks, Multi-task Learning, Object Detection, Classification, Defect Detection, Metal Gaskets

Download

IMEKO-TC10-2023-001.pdf

DOI

10.21014/tc10-2023.001

IMEKO TC

TC10 - Measurement for Diagnostics, Optimization and Control

Back to the proceedings