With the deepening of energy resource exploration in China, traditional lithology identification methods have become inadequate for the efficient and accurate analysis of massive core samples. The development of artificial intelligence technology has provided new insights for intelligent lithology identification. However, it still faces challenges under complex geological conditions, such as blurred lithological boundaries and insufficient multi-scale feature fusion. To address these issues, this paper proposes an innovative method that combines deep learning with fine-grained labeling technology. The study innovatively constructs residual fully connected networks (both serial and parallel), effectively solving the vanishing gradient problem in deep networks. Furthermore, it designs group-point and odd-even row-column sampling strategies. These strategies break through the traditional point-to-point prediction mode, significantly expanding the model's receptive field for regional lithological features. Additionally, a secondary fine-grained labeling system based on K-means clustering and expert knowledge is established, enabling the precise capture of complex geological features like transition zones and fractures. Experimental results show that using multi-dimensional RGB image features and a parallel network structure with a 19×19 sampling window, the model achieves an accuracy of 80% and a Macro-F1 score of 72.5% on the test set. This represents a significant improvement over baseline models. Notably, the recognition accuracy for monzogranitic gneiss exceeds 85%. This method maintains the model's lightweight nature while improving training efficiency by approximately 47%, demonstrating broad application prospects. It can be scaled for the automated interpretation of archived core samples to significantly reduce labor costs. Looking forward, it provides core technical support for deploying edge computing devices directly at exploration sites, driving geological exploration toward a real-time, intelligent paradigm and opening a new path for the deep integration of geoinformatics and artificial intelligence.