An On-line Ceramic Tile Classification System Using Adaptive Feature Selection

Theory in machine vision develops fast and there are more and more applications now, and as the kernel of artificial intelligence, the theory of machine learning also becomes more perfect. Combining machine vision with machine learning to make computers have learning ability becomes a trend in automatic and intelligent industry, and an on-line ceramic tile classification system is a typical example.In this paper, we tell the situations of artificial classification in ceramic tile industry in China, analyze the demand to an on-line ceramic tile classification system, improve the image pre-processing and feature selection algorithms considering some problems in applying, and propose an on-line ceramic tile classification algorithm using adaptive feature selection, which is now used in the factory after some tests and experiments.To solve the problem that there are no obvious difference in different kinds if ceramic tiles images, we improve the image gradient algorithm. The traditional image gradient algorithms are influenced by image noise and behave not so good, the improved algorithm gradient an image with variance to reduce the influence if noise and map it to logarithm domain to enhance textures of images.To solve the problem that there are only one or two training samples because of convenient using, we improve the feature selection algorithm. The traditional algorithms need more training samples to analyze and test, and cannot on-line select features with one or two training samples, the improved algorithm use a function to evaluate every single feature and estimate the degree of dispersal in every features using a priori knowledge after analyzing large number of image data when there is only one training sample.To solve the problem that some ceramic tiles are more similar than others and features of ceramic tile image change after a long time, we propose an on-line ceramic tile classification algorithm using adaptive feature selection. In the traditional multi-classification methods every sub-classifier has the same feature space, while in the proposed algorithm every sub-classifier has different feature space, also an incremental learning method is added in to solve the problem of changing of features.Experiments of different kinds of ceramic tiles in different time show, the accuracy of the system is nearly 0.0001 and the algorithm can also notice some unusual situations, including loss of corners and dirty traces in the surface of ceramic tiles, the innovations include:● We improve image gradient algorithm to reduce noise and enhance textures with variance gradient and nonlinear logarithm mapping.● We improve feature selection algorithm to on-line select features when there are only one or two training samples.● We propose an on-line ceramic tile classification algorithm using adaptive feature selection to select different features separately for every sub-classifier and to solve the problem of changing of features after a long time.