Answers are grounded in this paper and cite their sources.
With the rise of deep learning, detection methods based on it have made significant progress in chip surface defect detection. However, the performance in the detection of incomplete boundary defects in tiny chips is still unsatisfactory. Such defects typically present extremely weak features, fuzzy boundaries, low contrast, and easy loss during downsampling, leading to high missed detection rates, which has become one of the most challenging tasks in industrial defect detection. To this end, we propose an information compensation-based adaptive multi-scale fusion network (AMFNet) for chip small- and incomplete-boundary-defect detection; it consists of three core modules: graph attention-based feature extraction (GAT-FE), transformer-based feature information compensation (TFIC), and multi-level feature bidirectional fusion (MFBF). Specifically, the GAT-FE module enhances the representation of weak defect features by capturing long-range spatial dependencies. The TFIC module focuses on compensating for the lost detail information of incomplete boundary defects via global-context modeling. The MFBF module realizes bidirectional interaction and adaptive fusion of multi-level features, further improving the localization and classification accuracy of weak and boundary defects. Extensive experiments on a real-captured automotive MEMS industrial dataset and a public PCB dataset containing partial incomplete boundary defects demonstrate that AMFNet significantly outperforms state-of-the-art methods, especially in reducing missed detections of incomplete boundary defects and in model generalization. On the automotive MEMS pressure sensor dataset the network reaches 99.12% precision, 97.04% recall and 98.37% mAP. All reported results were obtained over five independent stratified re-partitions with distinct random seeds and are summarized as means ± standard deviations, together with 95% confidence intervals and two-sample t-tests, so that the reported gains can be assessed in terms of statistical reliability and reproducibility rather than by single-run comparisons.
No takes yet. Share an insight, caveat, or question.
Taiheng Liu (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: