Paper defect classification with neural networks
Modern web inspection systems have defect imaging capability. In addition to that, ABB Industrial Introduction Electro-optical inspection systems have been successfully used in paper mills for detection of various surface anomalies for more than thirty years. Their capability to identify e.g. holes, spots, edge cracks, and even wrinkles and coating streaks has been necessary for the mills in meeting customers’ quality requirements and in avoiding production disturbances e.g. due to web breaks or damages to soft calender rolls. Today, CCD (charge coupled device) cameras are dominant sensor types used in web inspection. The fairly recent developments have brought real-time imaging capabilities to them providing papermakers with new type of information about defects. The next step will be applying automatic image analysis. Some solutions in this direction and results achieved by ABB’s system are described in this article. Web inspection system Figure 1 shows the main parts of the ABB IndustrialIT Web Imaging ESI 7 system that has been developed by ABB Oy for inspection of paper web. The measuring frame consists of a sensor beam with intelligent line scan cameras and a light source beam. The number of cameras depends on the resolution required by the application. The high intensity light source is necessary for fast imaging and it also guarantees even illumination on the paper. The system adjusts light intensity automatically with varying paper grades, dust build-up, and lamp ageing.
A solid-state camera with linear CCD chip consisting of 512 … 4096 pixels is a natural choice for paper inspection [1]. Also matrix cameras are sometimes used, but they have several shortcomings in this application. Therefore, ESI 7 relies on the linear alternative. The light-sensitive cells (or pixels) of the linear chip generate electrical line images of the web with very short exposure times. The digitized images are combined by the camera processor to make up a complete, continuous image. The digital data stream coming from each camera can be tens of megabytes per second. The solution used in the ESI 7 to cope this huge flow of information is intelligent parallel processing in the camera to combine hole, spot, subtle defect, and streak detection into the same camera architecture.
 Figure 1 Web inspection system. Image analysis
Defects that were detected and previously classified as holes and spots can be now analyzed in more detail from the information available in the images. The steps needed in this pattern recognition process are shown in Figure 2.
 Figure 2 Steps in defect classification
Once the image acquisition is done by the camera, various pre-processing tasks are performed to make the image more suitable for the next steps. For example, dark level compensation and corrections of signal curvature and pixel-to-pixel variations can be done. A crucial, but also difficult, phase of the process is the image segmentation, i.e. separation of defective areas from their background. This is often done simply by thresholding the image and by drawing a rectangular bounding box around the defect, which means that a considerable percentage of the bounded area actually belongs to the background. It is obvious that this method provides only very approximate features. Therefore, ESI 7 is relying on more sophisticated, proprietary techniques that find the actual defect contours. Figure 3 shows a defect with box-segmentation and accurate segmentation lines.
 Figure 3 Paper defect segmentation The segmentation line and the area inside it are used for calculation of the features describing the defect. The features are chosen according to their capability to separate reliably defect classes from each other. The quality of the selected features is paramount, not the quantity of them. A high number of inconsistent features even tends to deteriorate the classifier performance. Defect classification
Traditional classifiers, e.g. the decision tree, rely on simple features such as size and shape measures. Because they are comprehensible, they are also easy to tune if the tree size is kept small. More abstract features, e.g. roundness and various grey level distributions, which are calculated from the image data, are necessary for classifying complicated defect types. It is very difficult to set manually limit values for these features because their meanings are not obvious. In addition, the number of these features may well be fifty or even more, which further makes the finding the right combination of values practically impossible. Then a neural network based classifier is a viable solution. MLP (Multilayer Perceptron) is one of the most well known neural networks. A simplified MLP is shown in Fig. 4. The basic idea with the MLP is that it is a nonlinear multivariate model that can be trained to approximate the dependence between defect features (input) and desired classes (output). The complexity of the network is selected according to the problem. In addition, decision tree can be used to divide the classification task into several branches on the basis of information, which is available from other sources than the image itself, each branch having a separate neural network.
 Figure 4 A simplified neural network Network training is done with a set of samples that have been selected and labelled (i.e. pre-classified) to be representatives of the defect classes. Labelling has to be done by an expert. It is advisable to include in each class several hundred samples - or even more if the variation inside the class is large. The whole task can be very laborious, because the database, from which the suitable representatives are picked up, often contains tens or hundreds of thousands of defect images. For this purpose ABB Oy has developed a special tool called Defect Viewer [2]. It is based on SOM (Self-Organizing Map), which is another type of neural networks, whose training is done unsupervised. The clustering algorithm helps to find various defect groups in the database and visualizes them to the user. In addition to that, Defect Viewer is an efficient tool for labelling defects according to desired defect classes. .
Figure 5 displays the performance figures of a neural network classifier at a paper mill. In this case about 7000 selected samples were divided into 18 classes. The average of the class performances was 87 %. These are typical results achievable when the first data set has been collected. The performance figures can be improved further by refining the training material and classes.
 Figure 5 Class performances Benefits The more precise classification provides several benefits for paper mills. For example, an off-line coating line is a critical process, where decisions about removing and patching defects at a rereeler or winder can now be done more reliably. Figure 3 shows a slime hole. The conventional system might classify the defect as a small hole that does not require any action. The MLP based classifier recognizes it as a slime hole whose total defective area is so large that it has to be patched. A reverse example is a large wet hole that actually has only a small insignificant hole inside and probably needs no action. Several new class names refer to defect causes, which helps and speeds up removing the actual origin of the problem. Examples are water spots, oil spots, precipitation spots and holes, contamination holes, wrinkles etc. At paper mills, knowledge about the right procedures regarding specific defects and removing their causes is accumulated by feeding class descriptions and recommended action instructions in the database. In addition, automatically generated statistics based on new detailed classes help to maintain the paper making process, because decisions can be made on the basis of facts, not of assumptions. References
1. J. Rauhamaa, Paper web inspection with intelligent line scan cameras. Machine Vision News, Vol. 6, 2001.
|
