<p>This study investigated the automated classification of different minerals utilizing three imaging techniques: thin-section microscopy, X-ray computed tomography (XCT), and smartphone-acquired digital photographs. To analyze these images, three convolutional neural network (CNN) models were employed: EfficientNet B3, ResNet 50, and GoogLeNet. A dataset comprising 13 minerals, including andalusite, andesine, anorthite, brucite, chlorite, cinnabar, diopside, erionite, oligoclase, quartz, tourmaline, tremolite, and triphylite, was systematically acquired and processed. A total of 10,764 images were acquired across the three modalities, consisting of 2,756 thin-section, 5,200 XCT, and 2,808 photographic images. Images were subjected to preprocessing methods, and partitioned into 70% for training, 10% for validation, and 20% for testing. Training was set to 60 epochs for all CNN models, to aid comparison. The results showed that all models yielded accuracy exceeding 95% for all image types. Comparative analyses revealed that XCT images and photographs performed slightly better than thin-section images. Although the accuracy and loss curves suggest that overfitting was avoided, the strong performance likely reflects the visually distinct characteristics of the mineral images, which enabled effective classification despite the small sample size. Still, the findings emphasize the importance of expanding the dataset to include multiple specimens per mineral and a broader range of mineral types. Additionally, insights from misclassified cases point to the need for standardized photography procedures and the integration of complementary physical and geological data. Overall, the study demonstrates the effectiveness of CNN approaches for efficient, accurate, and accessible mineral classification. It also highlights its potential in geological characterization, resource evaluation, and geotechnical engineering.</p>