A Comparative Study of CNN-Based Feature Extraction and Machine Learning Classifiers for Identification of Tyre Defect

Tyres play a crucial aspect in road safety and vehicle dynamics, as they provide an interface between the vehicle and the road, hence they are very susceptible to fatigue damage during service. In order not to further affect the service life of tyres and automobile safety, tyre manufacturing industries require efficient and accurate methods of detecting defects in tyres during the production process to ensure that defective tyres do not find their way to the market. In this study, we explore the performance of transfer learning using a pre-trained convolutional neural network (CNN) which includes ResNet-50 and VGG-19 for feature extraction, combined with other machine learning algorithms that include Random-forest, Logistic regression, and Support Vector Machine (SVM), for the final classification of the tyre defects. The performance of each of the combined CNN and traditional machine learning algorithms which include ResNet50 + LR, ResNet50 + RF, ResNet50 + SVM, VGG19 + LR, VGG19 + RF, and VGG19 + SVM, are evaluated and compared to other tyre defect classification result in which they combined the algorithm HOG + SVM, LBP + SVM, and HOG + LBP + SVM. The study concludes that VGG19 + LR and VGG19 + SVM methods show promising results for tyre defect detection as they optimise feature representation and classification accuracy.