Advanced Convolutional Neural Network Approach for Fabric Defect Detection

This paper introduces an innovative method for detecting defects in fabrics, employing a Convolutional Neural Network (CNN) optimized for binary classification. The distinctiveness of our approach is rooted in the utilization of a cyclical learning rate scheduler, enhancing the model's learning process over 100 training epochs. Crucially, both training and validation phases were conducted using an original dataset, meticulously collected and comprising three types of fabrics, each characterized by four distinct classes of defects. The CNN architecture, designed for this study, merges a detailed convolutional network for feature extraction with a dense layer, specifically calibrated for accurate defect detection in fabrics. The model's performance was rigorously evaluated against our unique dataset, yielding impressive results. An accuracy of 92% was achieved, signifying the model's effectiveness in defect identification across diverse fabric types. The recall rate stood at 91%, indicative of the model's proficiency in recognizing true defect instances. Precision was remarkably high at 98%, crucial for reducing false positives in quality control processes. The model also achieved an F1 score of 94%, balancing precision and recall effectively. The recorded loss function value of 0.2235 further attests to the model's capability in minimizing training errors. These outcomes underscore the potential of our CNN model, augmented with a cyclical learning rate, in transforming fabric defect detection, especially in settings that require handling varied fabric types and defect classes. The results presented in this study not only demonstrate the robustness of our model in a controlled environment but also suggest significant practical applications in automating quality control in the textile industry, elevating both accuracy and efficiency in defect detection processes.