Deep learning models for defect and anomaly detection on industrial surfaces

Abstract

Automated quality control is essential across various industries to reduce manual inspection and improve operational efficiency. While there are advances in computer vision and machine learning for defect detection, challenges persist, such as defect variability and the computational burden. This thesis presents specialized deep learning architectures addressing defect classification, segmentation, and detection in textiles, civil engineering, and manufacturing. For textiles, a novel system merges capsule networks with convolutional neural networks and a spatial attention module, achieving a 99.42% accuracy on the TILDA dataset. In civil engineering, the DepthCrackNet model, optimized for pavement crack detection, attains mIoU scores of 77.0% and 83.9% on the Crack500 and DeepCrack datasets. In manufacturing, the E-UNet3+ model for steel defect detection showcases a mIoU score of 86.19% on the SD-saliency-900 dataset. The research's core contribution lies in pioneering deep learning architectures that precisely detect defects across sectors.