Optimized Gabor Filter Banks and Autoencoder Models for Enhanced Knitted Fabric Defect Detection

Abstract

There is a high need for an automated system to detect fabric defects, as the current manual methods used in the garment industries in Nepal are unreliable and costly. Previous research has focused on specific fabric defects rather than overall fabric defects efficiently. This research employs two autoencoder models to identify different defects across different types of knitted fabrics, utilizing two datasets: the SFDG dataset and a custom dataset prepared from Butwal’s garment industries. The models benefit from a carefully designed Gabor filter bank to examine fabric compositions. This filter bank is fine-tuned by modifying parameters, wavelength, and orientation to detect varieties of defects in knitted fabrics. These models get feature representations from the Gabor filter bank’s outputs and help the system adapt to different types of defect patterns, making defect detection more reliable and accurate. The nearest neighbor density estimator finds possible defects and marks on the fabric images. The model's effectiveness and strength are shown by validating it on different types of knitted fabrics, including plain and patterned fabrics, using evaluation metrices like cTPR and ROC AUC. The first model achieves a cTPR of 0.879 and an AUC score of 0.947, while the second model achieves a cTPR of 0.899 and an AUC score of 0.958.