Research Of Color Steel Plate Surface Defect Detection And Classification Technology Based On Machine Vision

As a new type of material,color steel plate is widely used in construction engineering,home appliances and other fields due to its advantages of light weight,durability,beautiful environmental protection,thermal insulation and so on.At the same time,the surface quality of color steel plate is also an important indicator to evaluate its pros and cons.The surface quality problem not only affects the appearance,but also seriously reduces the product performance.How to quickly realize the detection and classification of surface defects of color steel plates,and further improve the production process has become an urgent problem for enterprises.In the process of automatic product inspection,machine vision inspection is advanced and efficient,which has become an indispensable core technology.In view of this,the project which is based on a color steel plate production enterprise,combing the research status at home and abroad and analyzing four common color steel plate surface defects including bruises,scratches,pits and indentations,researches the surface of color steel plate based on machine vision Defect detection and classification technology.The specific research contents include:(1)According to the relevant technical indicators of the enterprise,the design of the surface defect detection and classification system for color steel plates is completed from two parts of hardware and software.The hardware part mainly includes CCD industrial camera,optical lens,light source and illumination mode,and image acquisition card selection.The software part mainly includes the selection of Halcon10.0 image processing library and Visual Studio2015 development environment and the design of system software logic structure.(2)Aiming at the problem of noise on the surface defect image of color steel plate,an image filtering algorithm combining adaptive median value and inverse of gray level difference is proposed based on the analysis of noise type.The algorithm first compares the gray value of the pixel points in the neighborhood with the target point,determines the noise point type,and then classifies the noise points to achieve denoising.Through verification,the algorithm can effectively remove the noise in the image of the surface defect of the color steel plate,better protect the edge details of the defect,and have good adaptability to high-intensity noise.The histogram equalization is used to enhance the filtered defect image to make the defect contour clearer.(3)The different edge detection operators are studied,and the edge detection results of the four-color steel plate surface defect images are compared and analyzed.The results show that the Canny operator has the most accurate defect location and the best detection effect.(4)Considering the characteristics of surface defects of color steel plates,five-dimensional gray features,seven-dimensional geometric features and four-dimensional texture features are selected to construct 16-dimensional feature parameter vectors.In order to solve the problem that the eigenvalues extracted from the defect samples are quite different,this paper uses the unit interval method to normalize the data to improve the efficiency of defect recognition while ensuring the validity of the data.(5)As for the problem of traditional RBF(Radical Basis Function)neural network with poor anti-noise performance,slow training speed and difficult parameter determination,this paper proposes an improved genetic algorithm to optimize RBF neural network classifier.Firstly,based on the optimal individual preservation strategy of genetic algorithm,the selection operator is improved to improve the average fitness of the population and protect the diversity of the population.Secondly,the formula of crossover mutation rate is improved to improve the adaptability of cross mutation to solve the "premature" convergence of the algorithm.The problem is to improve the performance of the algorithm.Finally,it is the task to use the improved genetic algorithm to train and optimize the traditional RBF neural network to obtain reasonable network structure and parameters.Through verification,the classifier has high classification accuracy,fast recognition speed and small network training error,which is superior to traditional BP(Back Propagation)and RBF neural network,and has good application prospects.