Studies On The Key Techniques Of Defect Detection In Pulp And Paper Industry

Online inspection is an essential part of modern web or sheetmanufacturing. There are plenty of applications in different processes, e.g. inpaper, nonwoven, plastic, metal and plywood industries. The purpose of aninspection system is to detect and classify those defects that impair thequality of a product as compared to the requirements set by a user. Typicalcharacteristics of paper manufacturing are the large values of web width andproduction speed. The web width of a modern paper machine may exceed 9meters and its speed may reach 30 m/s. Such a machine makes about3X108 mm2 of paper each second, it is impossible to use human eyes toinspect the web. Inspection systems based on CCD technology adopt theCCD camera as the sensor unit, take a snapshot of running paper web, andtransmit the image to computer. Then with modern signal processingtechnologies, inspection systems can locate the position of paper defect andclassify the type of it.From the fewer reports of modern web inspection systems, one canconcludes that thresholding techniques(include global threshold and localthreshold) is the main algorithm of detection, while the self-organizing mapsand co-occurrence matrix are mainly utilized to extract features andclassification.This is mainly because the thresholding techniques have fastcomputation efficiency, they can satisfy the real-time demands of detection.But the self-organizing maps and co-occurrence matrix can detect thedefects more accurately, if their computation can be improved greatly, thenthey can be utilized directly to the detect the defects. This will overcome theweakness of thresholding techniques, such as the disability in detecting lowcontrast defects and weak defects.Based on the former algorithms of SOM, this paper proposes new ideasabout the key algorithms which influence the real-time computation efficiency.This paper also discusses the possibility of performing defect detection onGPU , and gets aspiring results. The term "real time" in this paper means 25frames/second or 30 frames/second, that is the speed of common homedigital video. Besides, there are many interferences in industry environment,such as chang of paper grade, dirty,etc. these directly influence themathematical model of detection. Normaly there are two ways of solving thisproblem. One is that the model can be updated momentarily to adopt thenew environment, the other is to develop new algorithm to overcome theinfluence of interferences.In section four a new method of is propsed whichcan update model momentarily.In section five a new defect detectionalgorithm based on singularity characterization is proposed, it can overcomethe interference. The main innovations of this paper are described as follows:1 Images of web can be considered as stochastic textures. This paperproposes new ideas about the key algorithms which influence the real-timecomputation efficiency. Firstly, a combination principle of confidence intervalsis proposed. It can merge nested confidence intervals, thus the computationload is reduced greatly. Secondly, the fast computation algorithms of twofeatures of co-occurrence matrix (mean and contrast) along vertical windowsthat are partly overlapped are derived. It makes it possible for the commonPC hardware to perform defect detection in real-time. With a cheaper CPU ofIntel P4 1.7Ghz, we get a speed of 5 frames/second on an image size of512X512.Due to the high speed of modern PC technology, there apear manynew techniques such as multithreading CPU, huge-cache CPU, dualcore/multi core CPU, PCI-EXPRESS bus,etc. These techniques can improvethe performance of CPU and the data transfer rates greatly. Along with thedevelopment of these techniques, we believe the proposed algorithm canperform the defect detection in real-time.2 High performance graphics processing units (GPUs) are the calculatingunit of modern graphics card. It plays the role of computation just like theCPU does in a computer. GPU has the parallel nature of computation. TheMoore's law of GPU is 6 months, which is 3 times than CPU's. All thesemake it possible for the algorithms depended on special hardware can beperformed by the CPU and GPU. In this paper the defect detection algorithmbased on statistical SOM is implemented in GPU. Also the algorithm isoptimized corresponding to the GPU's features. The algorithm has the abilityto perform defect detection in real-time on consumer market graphics. With acheaper GPU of Geforce FX5200, we get a speed of 8 frames/second on animage size of 512X512.The Geforce FX5200 has only 4 pixel pipelines, acore speed of 250Mhz, while newest graphics card, such as GeForce6800,has 16 pixel pipelines, a core speed of 1100Mhz, and supports thePCI-EXPRESS bus, so we can predict that with the development ofPCI-EXPRESS bus, the propsed algorithm can perform the defect detectionin real-time, that is 25-30frames/second.3 Learning is a very important procedure in the applications of artificialneural network;through it the weights of neural network will adapte to thesamples. The learning modes of SOM consist of incremental mode learningand batch mode learning. The batch-mode learning is completelydeterministic, but requires additional storage for each weight, which can beinconvenient in hardware applications. In this paper, with the powerfulstrength of modern GPU, the batch-mode learning process is implemented inGPU. The main precedures include computation of BMU and neighbourhoodfunction, sum of textures. Compared to the mothod based on CPU, thetraining time is reduced greatly, and the learning speed is improved. The fastcomputations of learning process have very wide applications in themanufacturing. It makes it possible for the SOM units to update online. Thereare many interferences in industry environment, such as chang of papergrade, dirty, etc., these directly influence the mathematical model ofdetection. If the SOM units can be updated quickly, then the mathematicalmodel of detection can be established in short time. This will reduce theproduction loss maximally.4 The wavelet transform has the ability to characterize signals both intime and frequency. Most of signal information is often carried by irregularstructures and transient phenomena. In many signals, interesting informationusually lines in sharp transitions. The web defects can be considered astransitions comparing to the background textures, they destroy thehomogeneous textured image. In this paper, a defect detection methodbased on the singularity characterization is proposed. Firstly, the originaldefect signal is convoluted with the smooth function, then a signal is selectedwhich both preserves the singularity of defect and removes small signalfluctuations. Then a wavelet transform is utilized with the selected signal.The most maxima lines corresponding to the stochastic textures are removed.Finally the intercepts of maxima lines are utilized to estimate the defect. Thealgorithm has fine features of anti-jamming. It can adapt the variation ofimage brightness, which makes it suitable for the industry environment.