ForgAnoNet: A Neural Network for Anomaly Detection in Artworks Using X-ray and Visible Spectrum Imaging

Forgery detection in paintings presents a significant challenge with substantial implications for the art world and forensic sciences. Given the high variability of artistic techniques and materials, forensic analysis must provide compelling, reproducible, and scientifically robust evidence. This paper introduces a novel technique for identifying anomalous regions in paintings, based on the detection of differences between visible and X-ray spectra, while also suppressing irrelevant artifacts, such as painting frames. Our model, the so-called ForgAnoNet, employs an architecture similar to O-Net but with several enhancements tailored to meet these specific needs. This architecture is the first to be applied to the fields of forensics and cultural heritage research. A methodology that is repeatable, accurate, and can suppress false detection from irrelevant irregularities. We proposed a novel neural network model that enhances both the precision and speed of detecting irregularities, such as cracks, voids, and previous restoration efforts. To evaluate the performance, we compared the methodology with five state-of-the-art models on the created datasets, which contained 4888 samples. A comprehensive evaluation of diverse X-ray images from various artworks demonstrates the effectiveness of our approach in practical applications. The newly developed ForgAnoNet achieves an accuracy of 98.08 %, significantly outperforming all other models in the study. Additionally, ForgAnoNet demonstrates precision, achieving a value of 0.4403, which effectively reduces false-positive rates and improves the reliability of anomaly detection in paintings.