Defects Localizations and Classification on Printed Circuit Board with Image Processing and Machine Learning

Siet, Joseph Jiun Wen (2023) Defects Localizations and Classification on Printed Circuit Board with Image Processing and Machine Learning. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Full Text -JosephSietJiunWen - RMH (1).pdf Restricted to Registered users only Download (18MB)

Abstract

In this current age of technological development, the evolutions of electronics, digital and computing technology has allowed the development of all kinds of programmable and nonprogrammable embedded devices which commonly have a compact and modular size though integrating various functions on it. In terms of this, Printed Circuit Board (PCB) has played an important role in integrating various types of functional circuitry including power circuitry, analogue circuits and digital logic circuits with minimal effort of space or path management in terms of circuitry wiring and components placement, since all the component footprints and conductive tracks have been well printed on it. However, during the process of handling the production of such PCB boards, errors such as scratches, speckles in copper plating and PCB features’ annotations may occurs, causing the formation of defects on the PCB traces. The examples of PCB defects included missing holes, mouse bites, open circuits, shorts, spurs and spurious coppers. These defects if left unnoticed on the PCB itself can cause device malfunctioning, destruction of circuitry components or even worse, harming the safety of the customers using the products. Hence, the quality control process before assembly of products is crucial to prevent selling any faulty products, where one of the most popular technology is knwon as Automated Optical Inspection (AOI) which detects and identify any defects present on PCB. However, the AOI process introduced by some big companies relies heavily on the usage of advanced machine vision hardware and high computational power, whichwould end-up costing a lot, and the solution becomes lacking in portability. In order to tackle such issue, this research project aimed for proposing a set of computer vision algorithms which helped to perform Automatic Optical Inspections (AOI) on the PCB, with the aids of portable lightweight edge computing device. The process of methods involved the image processing and defects collections through Computer Vision (CV) techniques, as well as the defects learning and classification through convolutional neural networks (CNNs), followed by implementation on the lightweight device such as Nvidia Jetson Nano and Raspberry Pi 4 Computer. The CV algorithms were implemented using OpenCV Python, whereas the CNNs algorithms were implemented using TensorFlow, KERAS package in Python language. The chosen CNN algorithms was expected to be the modified custom-built ResNet CNN model architecture, where it gave a test accuracy of 99.30% on the test PCB defects image datasets, which exceeded all the CNN model architectures of study in this research project that includes LeNet-5 and InceptionNet. Nvidia Jetson Nano Computer produced a frame rate of 72.67 FPS in ResNet architecture CNN model inferences based on TensorRT (TRT) based accelerated inference computations, which was higher than which produced by the Raspberry Pi 4 Computer (8.07 FPS), but with a trade-off of declined performance in model prediction accuracy while the system was implemented for the actual PCB inspection process.