A Cascaded Network for Surface Defect Detection on Lead Frames in Production Lines

Conducting surface defect detection on lead frames is a crucial operation in semiconductor manufacturing that significantly impacts product quality. This study develops a lead frame dataset and introduces a novel network tailored to detect minute target defects on lead frame surfaces across the production process. Initially, this study proposes a learnable hierarchical graph network (LHGNet) for effective feature extraction of small target defects, using learnable dynamic convolution modules for targeted feature extraction. Subsequently, this study introduces multiattention-based high-level screening-feature pyramid networks (MHS-FPN) and mixed cascade grouped attention (MCGA) modules, which enhance the feature expression of small targets while improving model classification performance. In addition, a loss function is specifically tailored for the lead frame scenario. Using the lead frame dataset and comparing it with competing methods, the feasibility and effectiveness of this model in detecting surface defects on lead frames in production lines are validated. The proposed surface defect detection you only look once (SDD-YOLO) network achieves a mean average precision (mAP) of 87.9% and an overall missed detection rate (OMD) of 2.06%, demonstrating its capability to either replace or assist manual inspections in assessing product quality.