Efficient Residual Learning With Multi-scale Attention for Robust Coal-rock Image Classification

Abstract Robust coal-rock image classification is challenging because visual differences among coal, gangue, limestone and sandstone are often subtle and can be obscured by lighting variation, dust, humidity, sensor noise and complex texture similarity in mining environments. This study proposes an efficient multi-scale attention residual learning framework for image classification, aiming to improve the balance among classification accuracy, parameter efficiency, convergence speed and cross-domain robustness. The proposed network introduces a residual convolutional block that integrates efficient multi-scale attention for spatial and channel-wise feature enhancement, while replacing conventional convolutions with depthwise over-parameterised convolution during training to improve convergence without increasing inference complexity. The model is evaluated on three public natural image benchmarks, CIFAR-10, CIFAR-100 and CINIC-10, and on a proprietary four-class coal-rock dataset. Experimental results show that the proposed method achieves competitive performance on standard benchmarks while maintaining a moderate parameter scale. On CIFAR-100, it improves classification accuracy over several attention-enhanced residual baselines, and on COALSet-4 it obtains a substantial improvement compared with the corresponding residual baseline, indicating stronger adaptability to fine-grained industrial visual patterns. Ablation experiments further suggest that both the multi-scale attention component and the modified convolutional operation contribute to accuracy gains and faster convergence. Visual inspection of attention regions indicates that the model can focus on discriminative local texture structures. The results demonstrate the potential of efficient multi-scale residual learning for coal-rock recognition and related industrial image classification tasks. Our source code is available at: https://doi.org/10.5281/zenodo.20841285}{10.5281/zenodo.20841285.