Enhancing image-based statistical process monitoring using Shearlet transforms and deep learning

With advancements in production processes and increasing product complexity, inspection operations now depend on numerous challenging measurements. Progress in image acquisition and machine vision has enabled capturing high-resolution images in short intervals, generating massive volumes of data. This study employs a combination of statistical process monitoring and deep learning to analyze image-based data. The process begins with feature extraction using the Shearlet transform, a powerful tool in image processing. A GLR control chart is then applied to evaluate process status, determining whether it is in-control or out-of-control. Performance of the proposed approach is evaluated using metrics including ARL, change-point estimation, and the Dice coefficient. Simulation studies demonstrate the superiority of the Shearlet-based method compared to competing approaches. Beyond detecting process status, three convolutional neural networks (CNNs), AlexNet, GoogLeNet, and ResNet50, are utilized to classify defect types. Transfer learning with ResNet50, pre-trained on ImageNet, assesses robustness against unseen defect classes through the Leave-One-Defect-Out (LODO) protocol. Results show all CNNs achieve high accuracy, with ResNet50 delivering the best performance and strong generalization under LODO evaluation. Finally, to explore transformer-based models for defect classification, the Data-efficient Image Transformer (DeiT-Small/16) is adopted, providing competitive accuracy and enhanced interpretability.