Self-Supervised Augmentation Strategies for Enhancing Painting Datasets in Low-Resources Scenarios

Abstract

Background: Painting classification is one example of how artificial intelligence (AI) is supporting artistic endeavors more and more, even in the absence of human intervention. Unfortunately, the lack of high-quality labeled data frequently hinders the effectiveness of AI-driven approaches, particularly in environments with limited resources. For deep learning models trained on an artistic set, these results overfitting and poor generalization ability are typical. Objective: The augmentation of the low-resource painting dataset using a Self-Supervised Learning (SSL) pre-training approach is examined and analyzed in this work, and its effectiveness for painting classification tasks is evaluated. Methods: The dataset used and collected from the WikiArt repository was carefully selected and contained 4,452 paintings. This dataset is used for the implementation of a quantitative experimental design. The research workflow generally consists of pre-processing, RandAugment integration, pre-trained SSL (SimCLR), and fine-tuning of ResNet18. Five model configurations were compared: ResNet18, ResNet18 + Pre-trained SSL (SimCLR), ResNet18 + RandAugment, ResNet18 + RandAugment + pre-trained SSL (SimCLR), ResNet34 + RandAugment + pre-trained SSL (SimCLR). The evaluation metrics used to measure model performance are accuracy, precision, recall, and F1 score. Results: The ResNet18 model trained with RandAugment and pre-trained SSL (SimCLR) produced the best classification performance in all experimental scenarios. The accuracy for this configuration was 78.41%. When compared to the standalone ResNet18, Pre-trained SSL (SimCLR), and augmentation configurations, the results show that augmentation diversity effectively supports SSL-based feature representation learning, resulting in improved generalization capability. Conclusion: To support SSL-based representation learning for the classification of artistic images under conditions of limited data, this study emphasizes enhancing augmentation diversity. The results show that in low-resource painting classification tasks, the integration of ResNet18+ RandAugment + pre-trained SSL (SimCLR) improves generalization capacity and reduces overfitting. Keywords: RandAugmentation, Pre-trained SSL, Data Augmentation, Classification, Painting, Low-Resource