Deep learning based rock type identification and mineral analysis for thin section rock images

Accurate, rapid, and reproducible interpretation of thin-section rock images is critical in geological research for petro-genetic interpretations, reservoir characterization, and engineering applications. Conventional thin-section analysis is highly dependent on expert judgment and is time-consuming, which motivates the need for automation. In this thesis, an integrated deep learning–based framework is proposed to perform rock-type classification and semantic segmentation of mineral phases from thin-section rock images acquired using a polarizing microscope. In the first stage, a dataset comprising 2,634 thin-section images representing three major rock groups (igneous, metamorphic, and sedimentary) was compiled. Transfer learning approaches based on VGG16 and EfficientNetV2-B0 were evaluated on this dataset, and a hybrid VGG16+EfficientNetV2-B0 model was designed by combining the feature extraction capabilities of both architectures. The proposed hybrid model achieved approximately 99% accuracy on the training and validation sets, demonstrating improved generalization compared with the individual models. In the second stage, a semi-automatic annotation workflow based on SLIC superpixels was developed to enable pixel-level labeling of thin-section microscope images, targeting the separation of quartz, alkali feldspar, and plagioclase phases using the generated ground-truth masks. Using this labeled dataset, three semantic segmentation approaches (U-Net, DeepLabV3+, and SegFormer) were trained under the same class definitions and using the same ground truth, allowing a consistent and comparable evaluation. During training, the accuracies obtained for U-Net, DeepLabV3+, and SegFormer were 94.83%, 95.90%, and 97.00%, respectively, while the corresponding mIoU values were 0.80, 0.86, and 0.87. On the validation set, U-Net achieved 88.10% accuracy with an mIoU of 0.6406; DeepLabV3+ achieved 90.45% accuracy with an mIoU of 0.66; and SegFormer achieved 94.56% accuracy with an mIoU of 0.73. These results indicate that SegFormer provides the best performance on the validation data in terms of both overall accuracy and mIoU, which reflects boundary quality, producing more consistent masks particularly in thin sections where grain boundaries are indistinct or phase transitions are gradual. Based on the resulting segmentation masks, modal proportions of quartz (Q), alkali feldspar (A), and plagioclase (P) were computed for each sample via pixel counting. These proportions were normalized