Textile Surface Defects Analysis with Explainable AI

The identification of surface level defects, such as those in manufacturing materials, is crucial in industrial processes. Deep learning methods have established their efficacy in defect detection. However, these black-box classification methods lack transparency, obscuring the rationale behind their classifications. This obscurity becomes a significant concern in safety-critical situations. Consequently, integrating explanatory mechanisms into these systems’ classification results is essential. This paper proposes a novel defect detection system based on Explainable Artificial Intelligence (XAI). A conventional Convolutional Neural Network (CNN) is employed to process an image database of fabrics, showcasing various manufacturing defects. This CNN, achieving a classification accuracy exceeding 93%, subsequently undergoes an interpretive analysis. To elucidate the CNN's output, a statistical method grounded in the SHAP (SHapley Additive exPlanations) value corresponding to a feature in the examined image is employed. Additionally, an alternative explanatory approach utilizing an unsupervised neural network is explored. This entails the use of Self-Organizing Maps (SOMs) to classify images in the dataset, facilitating visualization of the data categorization.