Lead frame surface defect detection based on geometric constraints and adaptive dual-branch normality recall for semiconductor manufacturing

Abstract

Surface defect detection in lead frames using machine vision is crucial for ensuring the quality and reliability of semiconductor manufacturing. While high-resolution industrial cameras are widely used, detecting tiny defects across hundreds of unit cells remains a significant challenge. Additionally, imbalanced defect data distributions in industrial environments exacerbate the detection challenge. To address these, a novel framework specifically designed for lead frame surface defect detection is proposed. The framework consists of two main components: multi-scale geometric constraint matching (MGCM) and dual-branch attention-based normality recall (DB-ANR). MGCM is a specialized algorithm designed for array-structured detection, efficiently extracting unit cells from high-resolution images by leveraging geometric constraints and eliminating redundant matching points, ensuring stable and precise results. DB-ANR addresses data imbalance and normal pattern forgetting by training on normal data to store patterns, which are recalled during inference to enhance detection accuracy. Additionally, the adaptive local–global dual attention module dynamically balances the contributions of local and global features, enabling robust detection across various defect types. Experiments on a self-constructed dataset with five types of lead frames demonstrate the effectiveness of the proposed framework. MGCM reliably extracts unit cells from high-resolution images, while DB-ANR achieves an AUC of 97.76%, PRO of 89.12%, and F1-score of 81.6%. The model also demonstrates efficient memory usage and fast inference speed, meeting deployment requirements in real industrial scenarios. Furthermore, the proposed array-oriented detection framework is not limited to lead frames and can be extended to other semiconductor applications, such as chip and wafer quality detection, where similar array-structured patterns exist.