A multiple-stage defect detection model by convolutional neural network

Abstract

Automated optical inspection (AOI) system is suffered from overkill and leakage problems. This study proposes a four-stage defect detection model, which uses convolution neural networks (CNNs) to examine product images for defect identification, classification, and positioning to reduce error rates and offer informative quality messages. An electronic component production line with two-camera CNN-based AOI system for appearance inspection is successfully modelled and experimented with promising performances. Stage 1 presents a CNN model of a binary classification for identifying good and defective products. Stage 2 presents a CNN model for categorial defect classification with multiple defects. Stage 3 further presents a CNN model to recognize multiple defects in a single product simultaneously and facilitate specification diversification. Finally, Stage 4 presents a CNN model for simultaneous multiple-defect classification and defect location positioning to provide advanced quality information. The proposed model achieves high precision and velocity as well as reduces labor power to facilitate quality assurance for a high-speed production line.