Recognizing the style of a fine-art painting with EfficientNet and Transfer learning

Due to the digitalization of the content of museums and art galleries in the last decades, a huge amount of digital fine art paintings became available on the internet. Therefore, the number of digitized paintings databases has increased rapidly, and it became difficult to manipulate their content manually. With the great performance of deep learning approaches and computer vision techniques, we became able to categorize paintings automatically. In this paper, we aim to investigate the effectiveness of the pre-trained EfficientNet models family for the task of identifying the style of a painting and propose custom models based on pre-trained EfficientNet models. We used transfer learning to fine-tune different types of pre-trained EfficientNet models from B0 to B6 on ImageNet. In addition, we added other layers to the base architectures of EfficientNet models to create our custom models. Finally, we analyzed the effect of retraining the last eight layers of our custom models. In our experiments, we used the standard fine art painting classification dataset Painting-91 with the same experimental setup. The deep retraining of the last eight layers of our custom models achieved the best performance with a difference of 5% than the base models.