A Lightweight Wafer Defect Pattern Recognition Network Based on Dynamic Snake Convolution

With the continuous advancement of semiconductor manufacturing processes, the morphological complexity and diversity of wafer map defects have significantly increased. This trend imposes more stringent demands on automated recognition systems in terms of structural modeling capabilities and classification accuracy. Existing deep learning models often struggle with elongated or curved structural defects due to limited receptive field flexibility and insufficient topological understanding, thereby constraining their practical applicability in industrial scenarios. To address these challenges, this paper proposes a lightweight wafer map classification framework that integrates dynamic snake convolution with a topological continuity loss. Extensive experiments conducted on the publicly available WM-811K wafer map dataset demonstrate that the proposed YOLOv8m-DSBlock-cls model achieves a Top-1 classification accuracy of 95.5%. The findings validate the effectiveness of incorporating structural modeling and topological supervision mechanisms, highlighting the framework's potential in enhancing recognition performance for structurally complex defects and offering a promising path for future research in defect-aware modeling.