To address the classification challenge of rock thin section images due to large intra-class variations and high inter-class similarities, a dual-path classification model WD-DDPM that integrates local frequency domain features and global semantics was proposed. Based on ResNet-50, a multi-scale downsampling module (WTMDB) based on wavelet transform to replace the traditional downsampling operation was designed. By performing wavelet decomposition on the image, it retains the low-frequency contour information while enhancing the edge and texture details in the high-frequency sub-bands using an attention mechanism, effectively strengthening the model’s ability to extract local discriminative features. Additionally, a U-Net encoder from the diffusion model was introduced as a parallel path to extract global context features containing overall structure and long-range dependencies. Finally, by fusing the dual-path features of local details and global semantics, a complementary image representation was formed and sent to the classifier for lithology identification. The experimental results on a dataset containing 108 rock thin slice samples show that the accuracy of the proposed method reaches 97.5%, the recall rate is 94.7%, and the F1 score is 96.0%, which is significantly better than mainstream convolutional neural network models such as VGG16 and MobileNetV3 (under the same experimental conditions), verifying its superior performance and practical value in the complex rock image classification task.