AUTHENTICITY DETECTION OF RADEN SALEH PAINTINGS USING SIMILARITY-BASED MACHINE LEARNING BY

• Main Author: Galih Mahar Putra, Benidictus
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78069
• Institution: Institut Teknologi Bandung

Raden Saleh is a pioneer painter using the first western style technique in Indonesia. One of the painting styles that Raden Saleh mastered was naturalism. Raden Saleh's paintings tend to be easily copied by painters with experienced skills. This is what counterfeiters use to create fake paintings. Experts have detected the authenticity of paintings by forensic tests or visually through the art criticism method. Forensic testing is expensive and damages the painting. The art criticism method assesses by describing the objects visible in the painting. This art criticism method is subjective and difficult to do for those who are new to the art world. Therefore, this final project builds a model that can detect the authenticity of paintings with quantitative objective assessment and provide a percentage of painting authenticity. Raden Saleh's works are hard to find in Indonesia. Painting authenticity detection research has used a classification approach. This approach has drawbacks due to the amount of data and retraining. The similarity approach is proposed to handle this problem. The model in the proposed similarity approach is Prototypical Networks with the Vision Transformer (ViT) encoder model of Segment Anything (SAM). Prototypical Networks predict by creating clusters from the data and calculating the distance of the cluster center point to the input data. This encoder model was chosen so that the model could learn the objects in the painting. The model is evaluated with cross validation and accuracy evaluation metrics as well as f0.5-score to minimize the cases of fake paintings predicted to be genuine by the model. The original painting data came from National Gallery of Indonesia and the Christie's Auction Hall website in Singapore. The fake painting data was created synthetically through Stable Diffusion because it is difficult to find fake Raden Saleh works in Indonesia. The prototypical network model produced 63% training accuracy, 66% validation accuracy, and f0.5-score 0.72 against cross validation and 66.66% test accuracy, and f0.5-score 0.71 against new painting test data. The important parameters in detecting the authenticity of paintings using Prototypical Networks are the capture angle and the learning rate.