A Computational Intelligence Framework for Artwork Image Classification Using CNN and ANFIS Models

The rapid growth of digital cultural heritage repositories has increased the demand for reliable and interpretable methods for artwork image classification. This study investigates explainable artificial intelligence approaches for fine-grained artwork classification by comparing a deep learning model and a neuro-fuzzy system in distinguishing paintings by Claude Monet from other Impressionist artists. A Convolutional Neural Network (CNN) architecture, SqueezeNet, is employed using transfer learning to automatically extract hierarchical image features, while an Adaptive Neuro-Fuzzy Inference System (ANFIS) is developed using twelve interpretable descriptors capturing color, texture, and structural characteristics of artworks. Both models are trained and evaluated on a dataset of 6000 digitized paintings under a consistent experimental framework. On the original dataset, SqueezeNet achieved higher classification accuracy (92.20%) than ANFIS (78.80%). However, evaluation on an independent dataset showed higher accuracy (75.00%) and F1-score (72.83%) for ANFIS than for SqueezeNet (62.58% and 68.97%, respectively). The results reveal distinct strengths in predictive performance and explainability. While SqueezeNet achieved higher predictive performance, ANFIS provided greater interpretability and demonstrated strong performance on independently collected data. These findings highlight the complementary potential of deep learning and neuro-fuzzy approaches and motivate further evaluation in multi-class artist attribution tasks.