Fine-Tuning Strategy for Re-Classification of False Call in Automated Optical Inspection Post Reflow

In the target classification and detection field, deep learning has recently demonstrated a lot of promise. An Automated Optical Inspection (AOI) consists of the vision and SMT components detection systems. The AOI machine scans the printed circuit board (PCB) through a vision system to detect defec...

Full description

Saved in:

Bibliographic Details
Published in:	2022 2nd International Conference on Emerging Smart Technologies and Applications (eSmarTA) pp. 1 - 5
Main Authors:	Jamal, IH, Nasir, MH A., Ani Adi Izhar, Che, Maruzuki, MIF, Ishak, KA
Format:	Conference Proceeding
Language:	English
Published:	IEEE 25-10-2022
Subjects:	Adaptation models Automated Optical Inspection (AOI) Convolutional Neural Network Deep learning Predictive models Pretrained Model Printed Circuit Board (PCB) Semiconductor device modeling Surface mount technology Surface Mount Technology (SMT) Training Transfer learning
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	In the target classification and detection field, deep learning has recently demonstrated a lot of promise. An Automated Optical Inspection (AOI) consists of the vision and SMT components detection systems. The AOI machine scans the printed circuit board (PCB) through a vision system to detect defects on the surface mount device (SMD) post reflow. However, there are times when the AOI machine produces verification calls, widely known as Calls. Currently, one widely method to reduce Calls is a two-step manual verification from an additional human operator. The human operator can lead to inconsistency throughout the inspection of Calls, whether a True Call or a False Call. This work assesses the application of deep learning in transfer learning and strategies to verify Calls from fine-tuning the pre-trained model. Since the Calls dataset is too small to train a deep model from scratch, we evaluated some transfer learning strategies from pre-trained models Xception. Since PCB images from the Calls dataset do not present the same complexity as the ImageNet dataset, feature extraction and fine-tuning from a less deep model as Xception has shown good results. The best-evaluated model obtained an accuracy of 91%.
DOI:	10.1109/eSmarTA56775.2022.9935366