open access

https://doi.org/10.14943/doctoral.k14145

https://hdl.handle.net/2115/78534

一様背景領域における欠陥検出のための凝集累積型明度遷移アルゴリズム

英語

Illumination invariance

2020-03-25

Hokkaido University

82p

Quality control is one of the most important procedure in modern industrial manufacturing. Product surface inspection is an important step in quality control to guarantee that the products show good impressions to consumers. In the past decades, product surface inspection was usually performed by human beings. Manually inspection is reliable for small number of products. However, it is impossible for human beings to inspect large number of products one-by-one. On one hand, it is too slow for human beings to do such kind of huge work. On the other hand, it is very expensive to hire workers for inspecting huge amount of products. Therefore, automatic vision inspection (AVI) is highly desirable. Conventional AVI algorithms can be classi ed into two categories: hand-crafted features based ones and data-driven features based ones. Hand-crafted features based approaches can be further classi ed into four categories: statistical methods, structure description and analysis, domain transformation, and data separation model based methods. Data-driven features based methods are generally based on deep learning techniques. These two kinds of algorithms have their own merits and demerits. Handcrafted features based methods usually have clear algorithmic meanings. For different application scenarios, suitable features can be designed for the unique goal. Hand-crafted features based methods are usually efficient, because they do not rely on learning from huge amount of data. Data-driven methods tend to design some learning parameters in a model and then train the model with data. The training data usually contains both of images and corresponding annotations manually marked by human beings. Although data-driven methods show high accuracy and high generalization ability, they need huge learning data and manual annotations. The training process also need huge computational resources and time. This dissertation focuses on detection of low-contrast defects in uniform background regions. To achieve this goal, an accumulated and aggregated shifting of intensity (AASI) has been proposed for detecting defects on micro 3D textured surfaces and an improved AASI approach with golden template picture (AASI-GTP) has been proposed for detecting defects on curved and highly-reective surfaces. About AASI, two novel features, named absolute intensity deviation and local intensity aggregation, which are associated with the probability of abnormality, are rstly designed to measure the saliency at pixel level. By considering the dynamic property during iteration and the salient features from initial image, AASI can iteratively shift the intensity of each pixel depending on its saliency, i.e., defect probability. The AASI output sequence along iterations of defective pixels follows an exponential function, while that of defect-free pixels can be formalized as a linear function. So the detection process can be regarded as a problem of tting to two statistical models. About AASI-GTP, a multiple-pairwise reective observation system is rst introduced. Multiple-pairwise cameras and diffused lights are xed in speci c positions and angles to capture all defects on the whole curved surfaces while avoiding shadows and highlights. Then a golden template picture (GTP) is introduced to represent the pixel-wise brightness of the defect-free background. By replacing mode intensity in AASI with GTP, the AASI-GTP is constructed for detecting defects on surfaces with uneven intensity. The overall dissertation is organized as follows. Chapter 1 introduces the importance of product quality control, the background of defect detection, conventional defect detection methods, challenges in defect detection, and contributions of this dissertation. Chapter 2 introduces the proposed accumulated and aggregated shifting of intensity (AASI) to solve the problem of detecting defects on micro 3D textured surfaces. Two salient features which measure the saliency in pixel-level are introduced. These two features are further used to construct an iterative pixel-level enhancement procedure, which can iteratively shift intensity of each pixel according with its defective probability. And then, an original tting-based classi cation rule was proposed. Two statistical models are utilized to judge whether each pixel in the given image is defective or not. Finally, the parameter setting was discussed. Chapter 3 introduces the improved AASI with golden template picture (AASI-GTP) for defect detection on curved and highly-reective surfaces. Firstly, the reection property of highly specular surfaces is analyzed. And then, a multiple-pairwise reective observation system is introduced. Multiple-pairwise cameras and diffused lights are utilized to capture all defects on the whole curved surfaces while avoiding shadows and highlights. Finally, the golden template picture (GTP) is introduced to replace the mode intensity in AASI, so that the AASI-GTP to be effective for surfaces with uneven intensity. Chapter 4 shows the dataset, experimental setting, experimental results and analyses. Experimental results demonstrate that our method generally outperforms state-of-theart unsupervised defect detection methods in terms of Precision and F-measure. AASI performs well even for low-contrast defects and small-sized defects with 3×3 pixels, and is robust to uniform illumination variations. And the AASI-GTP shows satis ed performance on curved and highly-reective surfaces. Chapter 5 concludes the whole dissertation and introduces future works.

2020-03-25

甲第14145号

博士(工学)

(主査) 教授 金子 俊一, 教授 山下 裕, 准教授 田中 孝之

情報科学研究科（システム情報科学専攻）

博士論文

Version of Record