Inception V3-Based Deep Learning Approach forCrack Detection in Paintings

This study proposes a deep learning-based methodology for the automatic detection of surface cracks in painting artworks, an essential task ‎in art restoration and preservation. A dataset comprising 700 high-resolution images (350 cracked, 350 non-cracked) was curated and expanded through data augmentation techniques to enhance model generalization. Preprocessing steps included resizing and normalization. Four deep ‎learning models, VGG16, ResNet50, a custom CNN, and InceptionV3, were evaluated. Among them, a fine-tuned InceptionV3 model ‎yielded the best performance, achieving a test accuracy of 98.01% and an F1 score of 0.98. Evaluation metrics such as confusion matrix, ‎accuracy/loss curves, and classification report validated the robustness of the proposed method. Compared to earlier approaches that were ‎restricted to structural domains or limited art-specific studies, the proposed framework demonstrates superior accuracy and broader applicability through transfer learning and advanced regularization. The system offers a scalable, non-invasive solution for digital inspection of ‎artworks and can be extended to diverse painting types, varied mediums, or severity-based crack classification. It also aligns with emerging ‎trends in lightweight, attention-guided deep learning models, supporting real-world deployment in cultural heritage preservation‎.