Unsupervised defect tetection in windshield glass via residual autoencoders and perceptual losses

Details

Supervisors

Delvenne, Jean-Charles

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en science des données, à finalité spécialisée: analyse de données
Master [120] : ingénieur civil en science des données, à finalité spécialisée: analyse de données

Abstract

enIn the glass manufacturing industry, detecting surface defects such as scratches, dents, and cracks is essential for ensuring product quality and structural reliability. Traditional inspection methods—both manual and automated—often struggle with small defect sizes, variable lighting, and the transparency of the material. This thesis proposes an almost fully unsupervised deep learning pipeline for defect detection in industrial windshield images, based on a residual encoder-decoder architecture trained with a custom weighted Structural Similarity (WSSIM) loss. The pipeline is divided into three stages: reconstruction, thresholding, and classification. High-resolution images from AGC's production line are preprocessed and split into patches, which are then reconstructed by the network to remove potential defects. Differences between original and reconstructed patches are scored using perceptual similarity metrics and passed through a thresholding mechanism to identify candidate defects. A lightweight classifier then distinguishes true defects from false positives and categorizes border artifacts. The encoder-decoder is trained without annotated defect data by leveraging defect-masked images and synthetic corruption.