MSAN-Net: An End-to-End Multi-Scale Attention Network for Universal Industrial Defect Detection

With the rapid advancement of automation and intelligence in the electronics manufacturing industry, the throughput of a single production line was grown exponentially. Although high efficiency brought significant cost and time advantages, it also led to two major challenges: <xref ref-type="disp-formula" rid="deqn1">(1)</xref> extremely low tolerance for error&#x2014;any slight defect might have caused the entire product to be scrapped; <xref ref-type="disp-formula" rid="deqn2">(2)</xref> increasingly diverse and more concealed types of defects&#x2014;bubble defects, internal chip defects, printed circuit board (PCB) defects, and specific process defects were continuously emerged, posing significant challenges to the inspection process. Traditional manual visual inspection or single-task deep learning models were often struggled to balance detection efficiency and accuracy in complex industrial scenarios. To address the above challenges, a single-stage industrial defect detection model based on multi-dataset mixed training&#x2014;MSAN-Net&#x2014;was proposed in this paper. Representative datasets covering the typical scenarios mentioned above were collected and organized, and part of the data was re-annotated to ensure a high level of consistency with actual production environments. MSAN-Net was adopted an integrated architecture, deeply combining UnifiedViT, C2f modules, convolution operations, SPPF structure, and Bi-Level Routing Attention mechanism to achieve accurate identification of complex industrial defects. Extensive experiments (including comparisons with multiple methods, ablation studies, and external validation) showed that MSAN-Net was outperformed existing SOTA models in industrial defect detection tasks, significantly improving detection accuracy for multi-class defects in complex scenarios, reducing reliance on manual inspection, and effectively lowering scrap losses caused by defects, thus providing a reliable solution for intelligent quality inspection in the electronics manufacturing industry.

Abstract licence: CC BY-NC-ND

18 more references, whose wording the paper’s licence keeps private.

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.