Flaw detection is important to quality assurance. The web inspection system as depicted in Fig. 1 has been widely applied by various industrials, such as plastic industry, paper industry, fabric industry, textile industry, and metalwork industry. This type of inline roll-to-roll inspection system has been successful in detecting defects. However, the function of discovering flaws by itself is not enough. It is desirable to identify the flaw types so that the system can immediately sort out the problem during the processing step and fix it immediately. Accordingly, we received 1676 defected images from one of the leading web inspection system developers in Taiwan to develop a methodology capable of identifying various flaw types of the flaw immediately after detection process. According to the defects in these images, they can be classified into 13 types, namely wrinkle, greasy stain, strip scratch, line scratch, dotted scratch, pinhole, protrusion, dent, bubbles, large stain, small stain, broken warp and broken woof. The representative images of the aforementioned 13 types of defects are shown in Fig. 2.
As a first step toward a successful flaw classification, it is crucial to correctly extract the desired flaws from the images. However, it is not trivial because these flaw images differ greatly. In particular, because flaws were collected from different roll-to-roll production lines, there were some variations in their backgrounds in terms of brightness and texture. Besides, the same type of flaws differs greatly in size, shape, intensity, and quantity. All these factors contribute to the difficulties in flaw detection. The most frequently used method for extracting flaw from images is image segmentation [1]. Image segmentation is a process to separate the desired flaws from the background. Zhang and Gerbrands [2] reported that there are more than 1000 different segmentation methods published by 1994. It is believed that this number might have doubled since then. Segmentation methods can be broadly classified as thresholding, clustering, edge-based methods, or region-based methods.
Thresholding [3][4][5][6][7] is the most popular techniques for classifying pixels into objects or background. Thresholding methods can further be classified as either global or local. A global segmentation method uses a fixed threshold to segment a whole image. On the contrary, a local segmentation method used multiple thresholds to segment an image region-by-region. The computational complexity of a global threshold technique is usually low. Segmentation methods that do not use spatial information to group pixels into regions are often called clustering [8,9]. The K-means algorithm [10,11] is one of the simplest clustering methods that iteratively partition an image into k clusters. Segmentation methods that use edge information to segment images are referred to as edge-based segmentation methods [12][13][14][15]. The basic idea of the methods is to detect edges and then link the discovered edges into closed curves representing meaningful object boundaries using edge-linking techniques. Generally, edge-based segmentation methods have the problem of instability. Region-based segmentation methods [16,17] refer to the methods that use region properties to divide an image into classes. The fundamental concept of these methods is to divide an image into regions of maximum homogeneity. Region-based segmentation methods include region growing [18], region