Abstract: To enhance the automation accuracy and efficiency of cylindrical roller surface defect recognition, a lightweight multi-view feature-driven classification model termed MFDANet is proposed. The model integrates a multi-view feature augmentation module to fuse raw texture, edge structure, and enhanced information for hierarchical feature extraction of subtle defects. A Haar wavelet downsampling strategy is adopted to reduce computational cost while preserving critical details. Moreover, an efficient channel-spatial attention mechanism is designed to adaptively emphasize defect-related features, and a dilated multi-scale feature extraction module is introduced to achieve collaborative modeling of local and global representations. Experimental results demonstrate that MFDANet achieves an accuracy of 87.70% and a macro-averaged F1-score of 87.33% on the cylindrical roller surface defect classification task. With only 4.41 MB of parameters and 0.026 GFlops of computation, the proposed model significantly reduces complexity while outperforming traditional classifiers, lightweight networks and recent advanced approaches, thereby verifying its efficiency and practicality.