Petrographic Identification from Small-Sample Thin-Section Images Using a Fine-Tuned Vision-Language Contrastive Learning Model

Accurate petrographic identification from thin-section images is important for geoscientific analysis but remains strongly dependent on expert interpretation. In this study, we organized 2634 polarized-light images from 324 distinct thin sections representing 108 lithologies and paired them with Chinese petrographic descriptions. Four conventional vision models were trained as closed-set lithology–classification baselines, and five CN_CLIP variants were fine-tuned for image–text alignment. To prevent overlap of images from the same thin section across outer folds, thin-section-wise three-fold evaluation was performed after excluding eight lithologies represented by fewer than three distinct thin sections; this protocol therefore comprised 2530 images from 311 thin sections and 100 eligible lithologies. At the fixed 20-epoch endpoint, CN_CLIP-ViT-L/14@336px and CN_CLIP-ViT-H/14 showed bidirectional image–text matching accuracies of 81.39 ± 1.97% and 81.31 ± 0.61%, respectively, whereas the conventional classifiers achieved Top-1 accuracies of 50.37%–58.20% under their separate closed-set task. These metrics correspond to different task formulations and are not directly comparable. Under the complementary image-level 8:1:1 protocol, CN_CLIP-ViT-H/14 achieved image-to-text and text-to-image matching accuracies of 93.75% and 93.38%, respectively, on the hold-out test subset. In an inference-stage ablation using the same checkpoint, hold-out test batches, and evaluation protocol, removing only the explicit lithology name reduced these accuracies to 82.72% and 77.21%, indicating that the rock-name token is informative but not the sole source of cross-modal discrimination. Without task-specific retraining, the same CN_CLIP framework also supported constrained lithology, major-mineral, and polarization-mode matching, with best task-specific accuracies of 90.57%, 66.51%, and 86.42%, respectively. Overall, the results support vision-language learning as a flexible framework for expert-assisted petrographic image–text matching on held-out thin sections within the available dataset, while emphasizing source-aware partitioning and cautious interpretation of task-specific accuracies.