Hokkaido University Collection of Scholarly and Academic Papers >
博士 （工学） >
Multiple Paired Pixel Consistency Model for Robust Defect Detection in Printed Logotypes
|Title: ||Multiple Paired Pixel Consistency Model for Robust Defect Detection in Printed Logotypes|
|Other Titles: ||印刷されたロゴタイプのロバストな欠陥検出のための複数ペアピクセル一貫性モデル|
|Authors: ||Xiang, Sheng Browse this author|
|Keywords: ||Mutiple paired pixel consistency (MPPC)|
|Position-indepent data inhibition Modification (PDI)|
|Issue Date: ||25-Mar-2021|
|Abstract: ||In modern manufacturing, quality control (QC) is an integral technique. Detection of product flaws plays a crucial role in quality management and ensuring that the product serves the consumer. In this study, we assess how currently available vision systems execute a range of QC tasks on printed items. In particular, we take into consideration the examination of printed characters/text or logotypes for defects, such as holes, scratches,cracks, and foreign artifacts. Printing-defect analysis is currently most done by human testers, and is labor-intensive and time-consuming job. Moreover, the results of an inspection might be unreliable because humans may arrive at different results depending on the time and the mood, skills, and experience of the inspectors. Therefore, human inspection is being replaced by automatic visual inspection systems.
Traditional defect-detection algorithms can be categorized as conventional feature-based and data-driven-based algorithms. Furthermore, conventional feature-based methods can be divided into four categories: statistical-, structure-, filter-, and model-based methods. Data-driven-based methods can be divided into two categories: traditional machine learning methods and deep learning techniques. Both data-driven-based and feature-based algorithms have their advantages and disadvantages. Feature-based methods usually have clear algorithmic implications and are therefore easy to control. Suitable features for various application scenarios can be configured for specific objectives. In comparison, feature-based methods are typically effective and simple to implement since they do not rely on massive quantities of data. However, certain difficulties do occur, such as the failure inability to detect small-sized defects and texture irregularities satisfactorily. Data-driven methods are usually implemented by designing certain learnable parameters of the model and then teaching the data model. Training data usually contains images and corresponding annotations that are manually annotated. While data-driven methods exhibit high precision and generalization, they involve a significant amount of learned data and manual annotations. The training phase often involves substantial computational capital and time.
This work focuses on the identification of printing defects on surfaces embossed using randomly spaced three-dimensional (3D) micro-textures. The embossing processes ∗Doctoral Thesis, Division of Systems Science and Informatics, Graduate School of Information Science and Technology, Hokkaido University, SSI-DT79185030, January 6, 2021. produces a very small variety of convex and concave shapes on the surfaces of metals, plastics, or other materials. Changes in the illumination on certain surfaces have a major effect on their appearance, resulting in difficulty in identifying defects. To realize this goal, the use of the multiple paired pixel consistency (MMPC) model was recommended. We first propose a consistency measure based on the correlation of consistent pixel pairs to obtain a robust defect-free model. We then set in motion a new assessment technique to accurately identify defects. Furthermore, a modification method called position-dependent data inhibition (PDI) is proposed to further improve the robustness and performance of the MPPC model.
This overall dissertation is structured as follows.
Chapter 1 introduces the importance of identification of defects and presents the associated works on defect detection. Some challenges in defect detection are discussed. Furthermore, the motivations and contributions of this study are explained.
Chapter 2 introduces the orientation codes (OCs), the use of which could reduce the influence of illumination fluctuations on defect detection; thus, OCs are used as the basis of the proposed method. First, we implement the original version of the OCs and then expand it by presenting two types of operations: a precise spatial differentiation for calculating the codes with a higher resolution and signed difference between any two codes as preparation for building up a more precise statistical model of their difference. Using these operations, we add a more reliable scheme to explain the statistical relationship between a pair of any logotypes pixels.
Chapter 3 introduces the proposed MPPC model in detail, including the fundamental principle and structure of the MPPC model. First, we observe the defect-free images of logotypes to determine the relationship between any pair of pixels in the logotype. We then introduce kurtosis to obtain the potential distributions. After analyzing the distributions, we explain how to pick the supporting pixel for each target pixel and finally construct the MPPC model for each pixel pair.
Chapter 4 provides a discussion of the method for using the proposed MPPC model of the relationship between pixel pairs in the defect-free logotype for identification of several types of logotype defects. Defect detection can be divided into two main stages. First, the status of each pixel pair is determined. We then identify the normal or abnormal states at the corresponding position defined by the elemental MPPC model.
Chapter 5 is focused on the PDI modification suggested on the basis of MPPC to further strengthen the robustness of the MPPC and stabilize its performance. The basic principles and mechanism of the PDI are explored in depth in this chapter. Finally, we verify the capacity of the PDI by appropriate experiments.
Chapter 6 lays out the experimental setup in detail. In this chapter, comparative experiments for the MPPC and MPPC+PDI models using several real defective images and synthesized or artificial defective images under different conditions such as illumination fluctuation and different noise intensities are designed. We test the robustness and efficiency of our methods through these experiments.
The final chapter outlines the key points of this study and presents a discussion of our algorithms. Finally, the strategy and idea for possible future works are discussed.|
|Conffering University: ||北海道大学|
|Degree Report Number: ||甲第14596号|
|Degree Level: ||博士|
|Degree Discipline: ||工学|
|Examination Committee Members: ||(主査) 教授 田中 孝之, 教授 小野里 雅彦, 教授 金井 理|
|Degree Affiliation: ||情報科学研究科(システム情報科学専攻)|
|Type: ||theses (doctoral)|
|Appears in Collections:||課程博士 (Doctorate by way of Advanced Course) > 情報科学院(Graduate School of Information Science and Technology)|
学位論文 (Theses) > 博士 （工学）