Classification of Paintings from Digital Images Using Feature-Driven EfficientNet Architecture

Traditional image classification models may not effectively differentiate between closely related categories, such as Paintings and Digital images (including cartoons, real-world photographs, computer graphics, and AI-generated images), which frequently intermingle and share similar color and shape features. Employing advanced feature extraction techniques, including HOG, color moments, LBP and GLCM features, we aimed to capture the distinctive characteristics of paintings. These features were processed and fed to a pre-trained EfficientNet-B4 model. Among these techniques, HOG features proved most effective results for distinguishing paintings from digital images. The EfficientNet-B4 model, with its compound scaling and high accuracy, was utilized to process these features and classify the images. Our approach achieved an impressive accuracy of 97%, demonstrating its effectiveness in distinguishing paintings from other digital images. This methodology not only advances the field of image classification but also enhances the efficiency and accuracy of digital art preservation efforts, providing valuable support for libraries, museums, researchers, and art enthusiasts.