<p>In the field of multi-light source illuminated defect classification, a defect’s visual properties such as shape, curvature, and depth vary depending on the illumination conditions, which are influenced by the direction of the lighting sources. These visual variations caused by illumination conditions can be key clues for defect inspection. These key features can be effectively extracted by predicting the variations in defects based on illumination conditions. However, conventional surface defect classification methods typically train separate models for each illumination condition and combine their results through an ensemble approach. These methods struggle to capture key features from variation across illumination conditions. To address these limitations, we propose a novel cross-reconstruction autoencoder, which transforms images captured under specific illumination conditions into images under other conditions. This autoencoder is trained to predict changes in defects based on illumination conditions, inferring intrinsic features of defects, such as their form and curvature with predictions under various lighting conditions. Additionally, we introduce a novel training strategy to optimize the autoencoder as a feature generator for defect classification. The effectiveness of the proposed method was validated using a publicly available multi-light-source illumination dataset. Experimental results demonstrate that our approach achieves superior classification accuracy, with improvements of up to 5.24% compared to conventional methods.</p>