Determination of various fabric defects using different machine learning techniques

Automatic systems are needed to recognize the fabric defects in textile manufacturing. For many years, systems that detect defects in the shortest time and with high accuracy have been researched. However, the multiplicity of the type of defect and the variability of the background made it difficult to recognize the defects. In this study, seven different feature sets based on Discrete Cosine Transform, Principal Component Analysis, Gray Level Co-occurrence Matrix, and Deep Learning Architectures have been used to detect defects. The extracted features have been classified by three different classification techniques (K-Nearest Neighbor, Support Vector Machine, and Decision Tree). The aim of this study is to investigate the effects of different feature extraction methods. The paper provides examination of the performances of different feature extraction methods and different classifiers. Methods have been evaluated in terms of precision, recall, F1-measure, and accuracy on two different datasets. mRMR feature selection after ResNet18 based feature extraction has enabled to obtain the features with the highest results in the study (0.831). Experimental results show that CNN-based feature extraction is more successful than statistical-based feature extraction methods.