Attention-Enhanced Deep U-Net for Chip Defect Segmentation: Performance and Limitations

In semiconductor manufacturing, reliable quality control means finding defects in very large, high-resolution micrographs that only affect a few pixels. This study reevaluates U-Net for chip-defect segmentation and methodically investigates the efficacy of attention mechanisms in enhancing performance. A deep U-Net with a ResNet-34 encoder is trained on 525 annotated 1024 1024 images of metal interconnects and vias (80/20 split). The training process employs Augmentations for data augmentation and a hybrid loss (Focal Generalized Dice) to mitigate extreme class imbalance. Three attention modules—Squeeze-and-Excitation (SE), Efficient Channel Attention (ECA), and CBAM—are integrated into the decoder and compared against a non-attention baseline. Performance is evaluated using Intersection-over-Union (IoU), Dice coefficient, background accuracy, and visual inspection of boundary continuity. All variants achieve strong results: the baseline reaches IoU 0.976, Dice 0.983, and background accuracy 0.995; CBAM slightly improves aggregate scores (IoU 0.978, Dice 0.985), while SE provides smoother edges. However, qualitative analysis indicates that ECA and CBAM often suppress low-contrast traces and cause breaks in thin, elongated interconnects; metric differences across models remain within 0.003 IoU. These results indicate that for structured, low-texture chip imagery with constrained data, a meticulously calibrated U-Net effectively captures critical cues, while generic attention modules yield minimal or detrimental impacts online continuity. Future research should focus on multiscale fusion and boundary-aware objectives instead of additional attention mechanisms and investigate transfer learning and domain adaptation to enhance applicability.