Accumulated and aggregated shifting of intensity for defect detection on micro 3D textured surfaces. (February 2020)

• Record Type: Journal Article
• Authors: Yan, Yaping; Kaneko, Shun'ichi; Asano, Hirokazu
• Is Part Of: Pattern recognition. Volume 98(2020:Feb.)
• Publication Date: 2020-02
• Subjects: Defect detection -- Accumulated and aggregated shifting of intensity (AASI) procedure -- Saliency description -- Illumination invariance; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.107057
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: This paper presents an unsupervised and training-free method for defect detection on micro 3D textured surfaces, which is difficult due to low-contrast and an unclear boundary between defect and irregular textured defect-free region. An accumulated and aggregated shifting of intensity (AASI) procedure is proposed to iteratively enhance defects, which solves the defect detection problem under a probabilistic saliency framework. A statistical distribution fitting rule is then proposed for pixel-level classification, which avoids the problem of collecting and labeling a large amount of data. Experiments on real-world industrial surfaces demonstrate the feasibility and robustness of our approach. Abstract: Micro three-dimensional (3D) textured surfaces are being designed for a lot of electronic products to improve appearance and user experience. Defects are, however, inevitably caused during industrial manufacture. They are difficult to be detected due to low contrast and unclear boundary between defect and irregular textured defect-free region. To achieve robust defect detection on micro 3D textured surfaces of industrial products, this paper proposes a probabilistic saliency framework with a novel feature enhancement mechanism. Two saliency features, absolute intensity deviation and local intensity aggregation, are designed to represent the pixel-level initial saliency. Based on these two features, an iterative framework, named accumulated and aggregated shifting of intensity (AASI), is proposed to shift the intensity of each pixel according to its saliency. Finally, all the pixels are classified as defective or defect-free by fitting the AASI iteration results to two statistical models, an exponential model and a linear model. Importantly, AASI procedure is unsupervised and training-free, so it does not rely on huge training data with time-consuming manual labels. Experimental results on a large-scale image dataset taken from real-world industrial product surfaces demonstrate that the proposed approach achieves state-of-the-art accuracy in industrial applications.