DeepGD3: Unknown-Aware Deep Generative/Discriminative Hybrid Defect Detector for PCB Soldering Inspection

Ching-Wen Ma, Yanwei Liu

Keywords: Unknown Awareness, Defect Detection, Generative/Discriminative Model

TL;DR: A generative/discriminative hybrid model effectively address the issue of performance degradation when the test samples come from new components for which no defective sample is available.

Abstract: We present a novel approach for detecting soldering defects in Printed Circuit Boards (PCBs) composed mainly of Surface Mount Technology (SMT) components, using advanced computer vision and deep learning techniques. The main challenge addressed is the detection of soldering defects in new components for which only samples of good soldering are available at the model training phase. To address this, we design a system composed of generative and discriminative models to leverage the knowledge gained from the soldering samples of old components to detect the soldering defects of new components. To meet industrial quality standards, we keep the leakage rate (i.e., miss detection rate) low by making the system "unknown-aware" with a low unknown rate. We evaluated the method on a real-world dataset from an electronics company. It significantly reduces the leakage rate from 1.827\% $\pm$ 3.063\% and 1.942\% $\pm$ 1.337\% to 0.063\% $\pm$ 0.075\% with an unknown rate of 3.706\% $\pm$ 2.270\% compared to the discriminative and generative approaches, respectively.