CNN Architecture on Distinguishing Art and Photo: A Comparison

Abstract

Painting and photography are growing with the advancement of technology and globalization. Painting and photography techniques are also developing along with the times. The art of painting is growing, marked by the increasing number of artists, especially in the realism genre. Many artists compete to create painting that is very similar to the original, making it difficult to tell the difference between a hand-drawn painting and a photo taken by a camera. Human limitations in distinguishing the two things need to be overcome with a system that is able to classify them automatically. We try to build a model for image classification, with the hope that the best model is able to classify which are photos and which are images. The best model is expected to be able to overcome human weaknesses in classifying which are photos and which are images. We compared three transfer learning architectures, namely the MobileNet-V1 architecture, the VGG-19 architecture, and the Xception architecture to find out which transfer learning architecture is the best for classifying the two classes, we choose several transfer learning architectural models that have been proven in previous studies to classify images with excellent result. After conducting and evaluating a series of experiments on each CNN model, the best model obtained for classifying images into painting or photo classes in this study is the Xception model trained with a dropout rate value of 0.5 which managed to gain a validation accuracy of 92.59% and test accuracy of 93.52%. The difference between the two values does not have much difference which indicates the model is not overfitting and also the other metrics score of the Xception model showing excellent result