Efficient Surface Defect Detection Based On Improved Otsu Segmentation And Saliency Analysis

Online detection of surface defects is an important part of the production process quality control, and the using of machine vision instead of human eyes for surface defect detection will become a trend. With the requirements for the production efficiency and product quality of modern manufacturing are continuously improved, the requirements for the technical performance of surface defect inspection system are increased, so it is necessary to study the theoretical problems related to lay a foundation for applications. Vision algorithms are the core technology of online visual surface defect detection system. A lot of research work has been done and some results are achieved, but the researches on vision algorithms are still needed in the following aspects: the enhancement of the adaptability and robustness of algorithms; using the images from online acquisition; solving the contradiction between efficiency and detection accuracy.Surface involved in the detection of surface defects can be classified into homogeneous surface and textured surface. For homogeneous surface defect detection, surface defect is easily to be segmented, so the method based on image segmentation is a common and effective method; but the complex situation in practical application may affect the result of image segmentation and defect detection, so the image segmentation algorithm under the real conditions is the key vision algorithm should be researched for this kind of application. For textured surface defect detection, the difference between the feature of defect and the feature of background cannot be simply described, and the feature of background is complex. Algorithms which can be applied to quickly and accurately detect defect objects from the images of different textured surface are key vision algorithms.To segment the surface image which is acquired under the real conditions, this paper improve the traditional Otsu algoritm which is widely used in the surface defect segmentation. The traditional 1-D Otsu algorithm has to exhaustively compute all between-class variances. Based on one characteristic of 1-D Otsu threshold, this work proposes a new fast algorithm. The new fast algorithm finds out every threshold which is equal to the integer part of the average of the mean levels of two classes, and then selects one threshold which is in accord with Otsu criterion.The analysis of 2-D Otsu algorithm shows that 2-D Otsu algorithm isn’t robust enough to noise. When 2-D histogram is segmented by 2-D Otsu threshold, within-class means are easily far from the main diagonal. Furthermore, this paper provides a new algorithm. The new algorithm gets a line intercept histogram directly from 2-D information of image. Then Otsu criterion can be used to find the best intercept threshold from the histogram. The new algorithm can avoid the disadvantage of 2-D Otsu algorithm. The new algorithm not only is faster than the fast algorithms of 2-D Otsu algorithms, but also gets the ablity of robust anti-noise.In the homogeneous surface defect detection, the contrast feature between object and background can be extracted simply after the segmentation based on the improved Otsu methods proposed, and surface defects can be determined by the threshold of the feature.In this paper, the texture surface defect detection method is proposed based on visual saliency analysis and fast Otsu segmentation. In new method, visual saliency map of the surface image is computed firstly, and then determine the presence of surface defect based on the saliency map. Because direct analysis for the specific features of surface defect is avoided and the visual saliency of surface defect is common in different applications, the new method is adaptive. This paper presents a model for computing visual saliency map, and the model has higher computational efficiency and accuracy than Itti, GBVS, DVA, AIM. Furthermore, this paper also proposes a method for determining surface defect based on fast Otsu segmentation and visual saliency map. The method firstly uses the fast Otsu algorithm proposed to segment saliency map and extracts the mean of background as the feature of saliency map, then implements the decision based on the feature. The computation of the method is efficiently and can meet the requirement of online detection.The accuracy of Otsu segmentation and saliency analysis will be reduced in the detection for small surface defect, thus it is necessary to study the small surface defect detection method. This paper provides a new method which can be used in the fast detection of small surface defects. Gradient method in spatial domain is applied to enhance the surface image. Try to detect the regions of small defects and improve the performance of Otsu method, the algorithm which is based on the distribution of variances of image blocks is developed to search the regions of small defects,and the algorithm can be successfully applied in the fast detection of small surface defects.In addition to the images used for comparing the algorithms proposed with other related algorithms, the data used in experiments includes the online acquired surface images of smart phone glass screen, touch screen mobile phone, and chemical fiber cloth. The experimental results verify the validity of the proposed algorithms in this paper.