A spatial-frequency information fusion network for mixed wafer defect recognition

Accurate wafer map defect recognition is critically important for the semiconductor industry. However, existing CNN-based defect detection models are limited in extracting contextually dependent features. In this paper, a spatial-frequency information fusion network that integrates a local feature extraction branch in the spatial domain with a global feature extraction branch in the frequency domain is presented for accurate mixed wafer map defect recognition. Specifically, a parallel spatial-frequency encoding backbone is built for synergistically enhancing the network’s capacity to characterise defect patterns. Meanwhile, aimed at the feature redundancy issue of existing CNN-based models, the proposed framework incorporates Octave convolution (OctConv) to design parallel spatial-frequency encoding backbone to achieve precise defect-focused capability. To align heterogeneous semantic relationships across domains, a Spatial-Frequency Feature Fusion module (SFFM) is introduced. It applies domain-specific attention to local spatial details and global frequency information, achieving progressive interactive fusion to boost encoder-stage feature extraction. Furthermore, an efficient Multi-level Cross-layer Fusion module, called MCFM, is presented. It uses projection to preserve complete information across encoder hierarchies, and realizes attention-guided bidirectional cross-layer interaction to integrate fine-grained details and semantic context for robust defect representations. Evaluations conducted on public MixedWM38 dataset demonstrate proposed model’s superior identification performance while validating the efficacy of each component. To promote open research, the complete codebase and model implementations have been made public at https://github.com/zzzzzqc/mix38_2.