A novel model for cultural-based classification of liberal arts using deep reinforcement learning.

Precise categorization of artistic styles, essential for art historical research, is hindered by the variety of cultural contexts that shape the appearance of artwork and the low precision of available methods. This paper introduces a new two-stage hierarchical model for robust cultural-based artwork classification using deep reinforcement learning for model optimization. This approach is driven by the hypothesis that decoupling cultural context identification from style recognition improves classification precision by enabling culturally-specialized analysis. First, a Convolutional Neural Network (CNN) identifies the cultural origin of an artwork (Western, Islamic, East Asian). Subsequently, style classification CNNs with different styles, customized for every cultural context and hyperparameters optimized using a novel deep reinforcement learning-based algorithm (LABHO), perform fine-grained style identification. The proposed model obtained 96.95% and 0.9581 accuracy and F-measure in cultural context identification, and 88.65% accuracy with 0.8439 F-measure in style classification. These results show a significant improvement in accuracy and efficiency over the conventional approaches, adding a more effective methodology for computational art analysis.