Research And Implementation Of Surface Defect Inspection System For Cold Rolled Strips Based On Neural Network

Quality control which has gained high significance in steel production and defects on the surface of cold steel strips was main factors to evaluate the quality of cold strips. Nowadays, steel companies generally adopt the behindhand manual visual method and strobe light detection for the quality of steel plate. However, these methods have many shortcomings, such as low inspection percentage, bad real-time performance, low detected degree of confidence, formidable circumstance of detection, etc. Therefore developing an automatic surface quality inspection system is a common demand for steel corporations. At present, artificial intelligence diagnostic system based on neural network is one of the hot items. It indicates its unique superiority in many aspects of pattern recognition, fault diagnosis, etc, because it has many characteristics such as massive parallel processing, distributed storage, learning capability and so on.This dissertation has thoroughly researched on related theories and key components of the cold steel strips surface quality inspection system and successfully applied neural network method into defect inspection of steel strips. The main content and achievement of this paper were listed as follows:The first is system designing. According to the cold steel strips surface quality inspection system's technical requirement, the paper put forward the system's design plan, established the system's workflow and the overall structure and carried out a detailed description of the structure of the system hardware and software processes; against the productive circumstance of steel plate, the author proposed relevant opinion on detection light source and CCD camera.The second is image processing. This article separately used the histogram equalization method and the image subtraction algorithm for defective image enhancement processing. The experiments indicated that the histogram equalization method not only can outstanding our interested defective region of image, but also can effectively restrain the effect of background noise; the feature of image subtraction algorithm was simpleness and practicality, also the effect of enhancement was right, but this algorithm was not suitable to deal with the image which contains complex background. At one aspect of image smooth processing, this paper adopted a mid-value filtering method which can largely improve the impaction of filtering, simultaneously by reserving most details of image. On the foundation of analyzing the insufficient of tradition Canny operator, the paper researched on edge detection method of steel surface image and realized the improved Canny operator edge examination algorithm. The experiment results showed that this algorithm can improve the accuracy and connectivity of edge orientation, restrain noise influence to some extent, and effectively increase the impaction of edge detection.The third is feature extraction. We extracted the feature of steel surface defective image and researched on some image features including state feature,grey-scale feature,textural feature and NMI feature altogether. We introduced NMI feature and invariant feature which meet geometrical invariance such as enlargement, translation and rotation. For complex defects on steel plate surface, both of features can provide effectively the basis for defect detection and recognition to some extent.The fourth is defect detection. This paper developed a system of quality detection in cold rolled steel surface and successfully applied neural network to defect detection. This paper has put forward a method of NMI feature combined with invariant feature to conceive the statistic feature of defective image. Simultaneously against the shortcomings of classical BP algorithm, we put forward a kind of BP improved algorithm with varying slope of activation function and dynamically adjusting different learning rate in order to accelerate the convergence of classical BP algorithm and to avoid plunging into the local minimum. The experiments showed that this BP improved algorithm had many merits such as high inspection speed, high discrimination and real-time capacity which can satisfy the demand of defect detection on steel plate surface, so it is an effective method. This system another innovation point lies in, in view of the limitation of the traditional defective image division algorithm, this system has realized a kind of real-time online inspection method with the free region of image. The experimental result proved that, compared with the traditional algorithm, the merits of this inspection method lie in it can adapt itself to the complex sample image analysis processing which some kinds of defects coexists and allow on-line inspection personnel to choose freely appropriate rectangle frame size according to the interested defective region of image, so this inspection method has strong flexibility, simultaneously the speed of examination was quick and completely satisfied the requirement of the real-time inspection, also it can provide the on-line inspection personnel the serious degree of defect analysis with rations.The last is defect classification and recognition. Considering the defect recognition problem, the paper also constructed a system defect classifier based on back-propagation network in this dissertation. This paper discussed features which were used in defect classification and increased Compactness feature,L–S factor feature and Linearity feature in the rectangular frame region as the basis of defect classification in time domain and simultaneously proposed a method which can extract features in the rectangular frame region of central bright area of defect spectrum image and increased mean and variance of gray value of all the pixels in this rectangular region as another important basis of defect classification in frequency domain. Through the experiments, it proved that these features were extremely effective to the cold rolled steel surface defect classification and this BP classifier had high identification accuracy and resolved the difficulty of defect classification on steel strips surface to some extent.