Study On Detection And Classification Of Cold Rolling Thin Strip Defects Based On Binocular Vision

At present machine vision detection technology is widely used in all walks of life because of its non-contact, high sensitivity and fast response, etc. Cold rolling thin strip are main products of metallurgy enterprises, which is used in car shell, national defense,building structure and the main raw materials of various kinds of electronic devices. In the pursuit of high yield, percent of pass and merit factor of cold rolled thin strip, the correct surface defect detection and classification have become a research hotspot nowadays.Combined with production practice, this article studies complex and varied forms of strip surface defects, complete defect segmentation and recognition in real time. The main research content is as follows:1. The real-time acquisition and analysis of hardware system are carried out. For high speed running of cold rolling thin strip, bright and dark field illumination configuration are used to realize detect without omissions. Besides, according to the actual testing environment, the cold rolling thin strip defect detection device is built.2. The defect of strip is identified in real time. High speed running of cold rolling thin strip leads to the amount of image data collected up to hundreds of megabytes, reduce the follow-up processing workload on time, conduct suspected defect judgment and save the defect images. Finally, its validity and accuracy are testified on the software system.3. The defect of strip is segmented. The paper’ aim is to analyze of the generation,harm and specific measures of specific defect withing noise and weak contrast of cold rolling thin strip production. Some work should be improved and innovated on the basis of predecessors’ threshold segmentation algorithm. By decomposing two-dimensional Otsu segmentation algorithm of two-dimensional histogram respectively, the threshold is calculated. In order to improve the oversight of edges and noise influence of noise in the process of decomposition, the balance factor is introduced on the pixel neighborhood grayscale histogram. The improved algorithm and quantum particle swarm double threshold segmentation algorithm and the traditional two-dimensional Otsu segmentation algorithm of segmentation are compared and analyzed respectively.4. The defect of strip is clustered and marked. The defect form after segmentation is analyzed. Considering the effect of scattered distribution of defects on late recognition classification number, the increased workload and decrease of classification accuracy. A new clustering merging algorithm is introduced, and the rectangular box is used to mark.5. The combined classifier which used to recognize defects is improved. For cold rolling thin strip surface defects, the analysis defect feature extraction and dimensionality reduction are verified, The traditional support vector machine of the small sample, nonlinearand high dimensional pattern recognition characteristics is applied to improve this situation.Besides, the particle swarm optimization algorithm is combined with cross validation to select the optimal parameters automatically. In order to solve the problem that a single classifier for defect sample dependence is so strong and the classification accuracy is low,the methods of combining the support vector machine(SVM), optimizing the BP neural network and probabilistic neural network are applied to identify and classify.