Research On Online Indentification Algorithm Of Float Glass Defect And System Implementation

The continue development of production level and technology of glass bring great challenges to the traditional manual detection methods. As a new inspection technology, machine vision with advantages of high speed, high precision and noncontact has adapted the modern production of glass. In order to modify the lag situation of detection algorithm study and system implement of machine vision on glass online detection in China, serial online detection and recognition algorithms of float glass are studied in this thesis, and some key technologies of float glass online detection system are realized.In the image preprocessing part, a simple and fast subtracting method is adopted to remove the complex stripe background. In order to segment the small and low contrast images, a segment method of twice-OTSU with 3σcriterion is proposed. With the help of OTSU, this method can derive the glass defect core precisely by selecting appropriate threshold automatically according to the different gray standard deviation of defect image, and so avoids the under-segmentation of low contrast images.After analysis of each type of glass defects, grid features, statistic features and multi-resolution energy features are studied respectively and extracted for the glass defects. Grid features reflect the local characteristics, energy features emphasize the entire property while statistic features have more pertinences with the glass defect. In the stage of grid features extraction, a self adapting border adjustment method is presented to make grid homogenization. In order to find the features with more relation with class, a sample unbiased evaluate method of SUReliefF is proposed based on the analysis of normal ReliefF. SUReliefF overcomes the ReliefF disadvantages, such as biased evaluation of features when the sample numbers of each class are imbalance, by adjusting the denominator weights and so makes the evaluation of features more fair and objective.ANN and SVM, which both have the similar structure, are studied comparatively for glass defect recognition. With the summary of factors which affect the generalization of ANN, BP net is improved by designing training algorithm and good sample selection. The SMO training method is used to improve the training efficiency and gets rid of the slow training and big memory requirement of normal SVM. The multi-class discrimination method of One-V-One makes the decision from SVMs more stability. An approach of glass defect identification based on multi-resolution analysis and information fusion(MRIF) combination is presented. The multi-resolution decomposition with 2D-DWT realizes the information hierarchy of defect image. Multi-classification decision on the approximated images realizes the information integrating and avoids the ones-idedness of single classification. A fusion method of HDC-DS based on evidence theory is proposed. This method can reflect the validity of the decision vector comprehensively by involving decision reliability, and overcome the bias of decision fusion of multi-classification.Finally, an online defect detection system of float glass is designed. The key issues of synchronous acquisition in detection process and plate synchronous of marking are studied, and application case is provided.