Abstract: The intelligent classification of metallographic microstructure images using machine learning methods represented by deep learning constitutes an important research area in materials informatics. In this study, a metallographic inspection image dataset for steels used in thermal power plants was established, and an improved ConvNeXt-based framework was developed for microstructure classification. The proposed model integrated pointwise convolution, Ghost modules and an attention mechanism, and it is further enhanced by data augmentation and transfer learning. Its performance was systematically evaluated through ablation and comparative experiments. The results show that the proposed framework achieves an mean average precision of 97%, a precision of 96%, and a recall of 96%, demonstrating its effectiveness for metallographic microstructure classification. To improve the model interpretability, a dual-level explanation framework was further constructed. At the local level, gradient-based saliency analysis was employed to quantify the contribution of feature channels and identify the regions most relevant to classification. At the global level, structured feature mapping, decision-tree modeling, and Sankey-diagram visualization was used to reveal the decision logic linking feature thresholds to microstructural categories. While verifying the classification accuracy of the model, this study improves the transparency and reliability of the model's decision-making process, and provides a technical solution with both favorable performance and interpretability for the intelligent classification of metallographic microstructures.