Rock image recognition based on deep learning serves as a critical task in intelligent geological analysis and core documentation. Although recent advances in deep learning have established the technical foundation for rock image identification, its practical deployment remains constrained by two main bottlenecks: intricate feature representations and ubiquitous label noise. Complexity of rock image features limits the effectiveness of intelligent model applications. Additionally, manual mislabeling and coarse marking issues during lithology annotation are often overlooked, leading to label noise in training datasets. To resolve these challenges, this study proposes a robust learning framework that integrates a novelty-driven feature extraction module with noise-resistant loss functions. Specifically, a Rock Hierarchical Heterogeneous Attention (RHHA) module is designed for enhancing rock feature extraction capabilities by applying distinct spatial and channel attention biases at shallow and deep layers of the network. Furthermore, Symmetric Cross Entropy (SCE) loss is used for loss calculation to enhance the robustness of models under noisy conditions. Comparative experiments were carried out on a real-world dataset of metamorphic rock images from northern Jiangsu Province, China. The results show that the proposed training framework achieves the best performance in rock image identification tasks, outperforming baseline and other comparative methods. The RHHA module outperforms other existing attention modules in capturing the texture and semantic features of rocks. In particular, the SCE loss effectively mitigates overfitting and maintains excellent generalization capability in the presence of noisy labels. The proposed framework holds promise for providing new insights into rock image recognition tasks.