A Comparative Evaluation of Rule-Based and Deep Learning Approaches for SAM Inspection of QFN Packages

Abstract

Scanning Acoustic Microscopy (SAM) is increasingly employed in the up-screening of commercial off-the-shelf (COTS) components for high-reliability applications, enabling detection of latent, field-failure-symptomatic, package-level defects beyond the reach of electrical test or visual inspection. As inspection volumes and acceptance stringency continue to grow, automated SAM image analysis is critical to sustaining throughput, consistency, and standards compliance. This work presents a comparative evaluation of two automated SAM inspection paradigms applied to Quad Flat No-Lead (QFN) packages: classical rule-based machine vision and supervised deep-learning inspection. Both approaches were assessed on identical top side and bottom side SAM datasets against industry-standard-informed customer requirements, using classification metrics and region-level spatial agreement relevant to quantitative accept/reject criteria. Deep learning models achieved strong classification performance for complex top side defect morphologies, while rule-based inspection demonstrated superior accuracy for bottom side die-attach voiding and perimeter delamination, where precise area-based thresholds govern acceptance. Limitations in region-level quantification, sensitivity to scan variability, and near-threshold classification were observed for the deep learning approach. These findings indicate that a hybrid strategy—combining deep learning for complex defect detection with rule-based analysis for standards-driven quantification—offers the most robust and defensible framework for automated SAM screening in high-reliability QFN up-screening workflows.