Automated surface defect detection framework using machine vision and convolutional neural networks

• Published in: Journal of intelligent manufacturing Vol. 34; no. 4; pp. 1995 - 2011
• Main Authors: Singh, Swarit Anand, Desai, K. A.
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.04.2023; Springer Nature B.V
• Subjects: Artificial neural networks; Automation; Business and Management; Centerless grinding; Control; Data augmentation; Datasets; Deep learning; Feature extraction; Image classification; Inspection; Machine learning; Machine vision; Machines; Manufacturing; Manufacturing industry; Mechatronics; Neural networks; Processes; Production; Quality control; Robotics; Rollers; Small business; Support vector machines; Surface defects; Training; Vision systems; Deep learning; Surface defect detection; Machine vision; ResNet-101; Centerless grinding
• ISSN: 0956-5515; 1572-8145
• DOI: 10.1007/s10845-021-01878-w

Machine vision-based inspection technologies are gaining considerable importance for automated monitoring and quality control of manufactured products in recent years due to the advent of Industry 4.0. The involvement of advanced deep learning methods is a significant factor contributing to the advent of robust vision-based solutions for improving inspection accuracy at a significantly lower cost in manufacturing industries. The requirement of computational resources and large training datasets hinders the deployment of these solutions to manufacturing shop floors. The present research work develops an image-based framework considering pre-trained Convolutional Neural Network (CNN), ResNet-101 to detect surface defects with the minimum training datasets and computational requirements. The outcomes of the proposed framework are substantiated through a case study of detecting commonly observed surface defects during the centerless grinding of tapered rollers. The image datasets consisting of standard tapered rollers and three common defect classes are captured and enriched further with the help of the data augmentation technique. The present work employs ResNet-101 for feature extraction combined with and multi-class Support Vector Machine (SVM) as a classifier to detect defective images. The effects of the feature extraction layer ( fc1000 ) and pooling layer ( pool5 ) activation are explored to achieve the desired prediction abilities. The testing trials demonstrate that the proposed framework effectively performs image classification, achieving 100% precision for the ‘Good’ class components. The study showed that the proposed approach could overcome the requirements of large training datasets and higher computational power for deep learning models. The proposed system can be of significant importance for Micro, Small, and Medium Enterprises (MSMEs) and Small and Medium-sized Enterprises (SMEs) as an alternative to conventional labor-intensive manual inspection techniques.