Impact of Image Augmentation on Deep Learning-Based Classification of Granite Tiles

Abstract

Automated image classification of granite tiles presents challenges due to the variable textures present in the images. This study explores the impact of image augmentation on the classification performance of granite tile images using the pre-trained VGG16 Convolutional Neural Network (CNN) architecture. Contrary to the expected benefits of image augmentation in enhancing model robustness and generalization, our experiments highlight potential drawbacks when applied to texture classification. Using transfer learning, the baseline model, trained without augmentation, achieved the highest accuracy of 97.1%. However, most models trained with augmented data exhibited reduced accuracies, with the exception of flipping technique reaching up to 98.5%. Notably, common augmentation techniques, including rotation and scaling, demonstrated varying degrees of performance degradation, suggesting that these techniques may distort essential textural patterns crucial for accurate classification. These findings underscore the importance of employing dataset-specific augmentation strategies tailored to the unique characteristics of the dataset, particularly in texture-sensitive applications like granite tile classification. © 2024 IEEE.