View in EDS

A lightweight high precision real-time system for bonding ball–pad localization in industrial packaging inspection.

Saved in:
Bibliographic Details
Title: A lightweight high precision real-time system for bonding ball–pad localization in industrial packaging inspection.
Authors: Wei, Haoran1 (AUTHOR) 2020241835@jsnu.edu.cn, He, Zhenzhi1 (AUTHOR) hezz82@163.com, Ye, Guo1 (AUTHOR) yeguo5818@126.com, Shu, Xiaotong1 (AUTHOR) xiaotongshu98@gmail.com, Peng, Jingshong1 (AUTHOR) PengJinsong_jsicat@163.com, Chen, Tianchi1 (AUTHOR) ctc900112@163.com, Sheng, Lianchao1 (AUTHOR) shenglianchao@163.com, Lu, Xiangning1 (AUTHOR) lxnam89@163.com
Source: Measurement Science & Technology. 2026, Vol. 37 Issue 27, p1-17. 17p.
Subjects: Wire bonding (Electronic packaging), Real-time computing, Artificial neural networks, Inspection & review, Electronic packaging
Abstract: As a critical interconnection technology in chip packaging, wire bonding imposes stringent requirements on the localization accuracy of bonding balls and pads. Conventional vision-based inspection methods often struggle to satisfy real-time constraints when targets are extremely small, densely distributed, and embedded in cluttered, reflective industrial backgrounds. To meet these demands, a real-time high-precision localization approach is proposed based on an improved real-time detection transformer framework, termed Dynamic-Fast-Aware DETR (DFA-DETR). DFA-DETR integrates several task-oriented architectural components to improve feature representation, efficiency, and robustness under bonding ball–pad detection scenarios. Specifically, inspired by FasterNet and ResNet, the backbone is constructed as a lightweight ResNet-FasterBlock network, which uses partial-channel spatial convolution to reduce redundant computation while preserving feature representation capability for tiny targets. Moreover, a SlimNeck structure is employed for efficient feature fusion and further optimized with DySample to replace fixed upsampling, alleviating spatial misalignment across scales for small bonding ball–pad objects. In addition, a magnitude-aware attention module (magnitude-aware linear attention-adaptive intra-scale feature interaction) based on linear attention is incorporated to enhance region discrimination and suppress strong background interference typical in industrial imaging. Experiments on the bonding ball–pad positioning task achieve 77.92% mAP50–95 and 90.47% mAP75, together with an overall F 1-score of 96.62%. The resulting model runs at 86.81 frames per second with 16.30 m parameters, indicating strong potential for real-time industrial deployment. Ablation studies and qualitative analyses further confirm the effectiveness of each component and their complementary effects. Overall, this work provides an accurate and efficient solution for online visual inspection in packaging wire-bonding processes. [ABSTRACT FROM AUTHOR]
Copyright of Measurement Science & Technology is the property of IOP Publishing and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:As a critical interconnection technology in chip packaging, wire bonding imposes stringent requirements on the localization accuracy of bonding balls and pads. Conventional vision-based inspection methods often struggle to satisfy real-time constraints when targets are extremely small, densely distributed, and embedded in cluttered, reflective industrial backgrounds. To meet these demands, a real-time high-precision localization approach is proposed based on an improved real-time detection transformer framework, termed Dynamic-Fast-Aware DETR (DFA-DETR). DFA-DETR integrates several task-oriented architectural components to improve feature representation, efficiency, and robustness under bonding ball–pad detection scenarios. Specifically, inspired by FasterNet and ResNet, the backbone is constructed as a lightweight ResNet-FasterBlock network, which uses partial-channel spatial convolution to reduce redundant computation while preserving feature representation capability for tiny targets. Moreover, a SlimNeck structure is employed for efficient feature fusion and further optimized with DySample to replace fixed upsampling, alleviating spatial misalignment across scales for small bonding ball–pad objects. In addition, a magnitude-aware attention module (magnitude-aware linear attention-adaptive intra-scale feature interaction) based on linear attention is incorporated to enhance region discrimination and suppress strong background interference typical in industrial imaging. Experiments on the bonding ball–pad positioning task achieve 77.92% mAP50–95 and 90.47% mAP75, together with an overall F 1-score of 96.62%. The resulting model runs at 86.81 frames per second with 16.30 m parameters, indicating strong potential for real-time industrial deployment. Ablation studies and qualitative analyses further confirm the effectiveness of each component and their complementary effects. Overall, this work provides an accurate and efficient solution for online visual inspection in packaging wire-bonding processes. [ABSTRACT FROM AUTHOR]
ISSN:09570233
DOI:10.1088/1361-6501/ae7f3f