A Deep Learning Approach Based on Image Patch Sets for Art Forgery Detection

With the recent expansion of the art auction market, the discernment of forged artworks has become increasingly vital. Studies have been conducted to detect forgeries through various means, such as physical examinations of paint and canvas, as well as more abstract inquiries into the artistic style. Among these, style-based studies have faced challenges due to the lack of relevant datasets. To address this, we have constructed a dataset by manually creating both genuine and forged oil paintings. Typically, artwork images are very large. Previous research has extracted small image patches for input but often failed to represent the artwork’s features, depending on the patch’s location within the work. In this paper, we propose a deep learning approach that utilizes a set of patches instead of a single image patch to determine whether the given artworks are from the same artist. Using multiple image patches is advantageous because they can represent the characteristics of the artwork more effectively than a single image patch. Experimental results demonstrate that the proposed approach achieves an accuracy ranging from 76% to 99%, with the accuracy increasing as the image patch set size grows.