Improving machine vision crack die detection with multi-kernel learning SVM

Abdullah, Mohd Khairi John (2024) Improving machine vision crack die detection with multi-kernel learning SVM. PhD thesis, University of Nottingham.

[img] PDF (Thesis - as examined) - Repository staff only until 8 March 2026. Subsequently available to Anyone - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Available under Licence Creative Commons Attribution.
Download (27MB)


With the rising need to produce semiconductor parts with zero defects to support the growth of the automotive industry, where safety must be assured, it is vital to perform visual inspection throughout the assembly process to isolate and detect defects. Visual inspections performed by a human have two generic issues. Firstly, the variation in the measurement taken due to variations in readings taken by different individuals would impact the decision to separate and remove the defective products. Secondly, the inaccurate measurement of the size and decision makes it difficult to analyse.

Automated machine vision inspection is generally the answer to the manual inspection problems, as it is consistent, and all measurements, images and decisions are captured and processed automatically. Unfortunately, the system can have less accurate detection capabilities in comparison to trained humans when there is a minimal difference due to a slight difference in contrast or when the defects do not have a consistent pattern or characteristic. Such shortfall would take longer processing time to provide the result or fail to provide a result, thus making manual inspection more effective.

One of the defects in this category is the cracks on the die. Reasons that make cracks hard to detect; firstly, the slight difference in contrast (intensity) between the crack and the background. Secondly, the characteristic of shape is inconsistent. Thirdly, there are very few samples of such detects as it seldom happens. Fourthly, the crack's size is as small as 3 microns (1 micron = 0.001 mm). Fifthly, to capture, the system should be sensitive and have high ranges of contrast ratio. An offline vision system with a high-power scope can detect cracks of less than 9 microns, and an inline vision system can usually detect cracks of more or equal to 9 microns only.

In this study, the project aims to improve the inline machine vision system to detect cracks on the semiconductor die by maintaining the current automated vision inspection system and improving defect detection using machine learning. The expected outcome is an inline machine vision system capable of effectively detecting cracks 9 microns in length, width, and radius.

Firstly, to analyse the shortcomings of the current methods, a review was carried out. The finding was that the current solution does not mimic the human eye that detects cracks based on the slight contrast change. Secondly, to identify an effective method of capturing an image, several image capturing methods were explored. From the observation, it was found that different lighting helps to improve the image quality for a specific situation. For this research, we conclude that red LED lighting on the inline inspection machine was sufficient. Next, to develop an effective machine learning method for crack detection, several methods were studied. From the observation, we found that the slight difference of the crack from its surface is linear; thus, Support Vector Machine (SVM) LinearSCV kernel was used. Data extraction was conducted using edge detection methods such as Gabor, Canny and Roberts. It was fed to LinearSVC, resulting in an effective local area thresholding for those pixels with slight differences, which gave a result of 75% to 77% pixel-by-pixel accuracy and the ability to detect the crack on the images by 100%. To further improve the method, an effective artificial intelligence (AI) method such as SVM with multiple kernel learning (MKL) was proposed. Since the effectiveness of the local thresholding meets the expectation, this kernel and the data extraction method were maintained for MKL implementation. Five SVM RBF kernels were used along with a LinearSVC kernel to create an MKL kernel that increases the crack pixel's contrast and masks the rest of the pixels that give the result needed. Pixel-to-pixel accuracy of 99.6% to 99.8% was achieved, with 100% accurately detecting the crack on the images. The invention is generic; with minor tweaks, it can be implemented to detect cracks on different surfaces.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: Nugroho, Hermawan
Mohd Isa, Dino Amshah
Keywords: machine vision inspection, semiconductor die, crack detection, machine learning, Support Vector Machine (SVM)
Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering
Faculties/Schools: University of Nottingham, Malaysia > Faculty of Science and Engineering — Engineering > Department of Electrical and Electronic Engineering
Item ID: 76585
Depositing User: ABDULLAH, Mohd
Date Deposited: 09 Mar 2024 04:40
Last Modified: 09 Mar 2024 04:40

Actions (Archive Staff Only)

Edit View Edit View