Brushstroke analysis for oil painting authentication based on image-to-image translation with generative adversarial networks

Abstract

From an art history point of view, oil painting authentication evaluation was always a crucial matter. Recently by advancing computer vision and image processing methods, many studies have been done for detecting oil painting forgeries. Although Computer-based methods such as Deep Learning and Machine Learning illustrated promising results, more research should be done to design the AI model to detect the artist's unique brushstroke style for detecting forgeries. Since the artist’s unique style is one of the key criteria for distinguishing original from fake artworks. We hypothesize that using image-to-image translation methods such as Conditional Generative Adversarial Networks (CGANs) we are able to generate the artist's unique brushstroke style in order to evaluate paintings authentication. The purpose of this research is to develop an automated system to analyze brushstroke for oil painting authentication using Conditional Generative Adversarial Networks (CGANs) and dissimilarity measurement. We design a specific CGANs for each style of the artist (aimed painter) to generate the painter’s style pattern. Next, in order to examine the suitability of CGANs’ structures for learning and generating trained painter’s brushstroke style, we experiment with three different discriminators. In this research, to make painting’s brushstroke dataset, we use Reflectance Transformation Imaging (RTI) method, which is a computational photographic technique that captures a subject’s surface shape and color and enables the interactive re-lighting of the subject from any direction. We examine the system using three different states of given new paintings. In our experiment the new painting has three different states including paintings drawn by the aimed painter (real), painting from another painter with close brushstroke style (fake) and in third state artificial forgery painting based on aimed painter painting (made by us due to lack of forgery dataset). After that, using Fourier Transform we measure the dissimilarity between Gan-Generated RTI brushstroke image of new painting and RTI ground truth brushstroke image to determine the probability of being fake and real painting. The results of dissimilarity measurement value show the promising results of using the system for detecting the probability of being fake and real painting

however, the results for artificial forgery painting is so close to the real painting. Comparing the results of Patch-Gan discriminator with Image-Gan and Pixel-Gan discriminator by dissimilarity measurement shows that the Patch-Gan discriminator is a more suitable GANs design for being used in oil paintings authentication.