Research On Object Detection And Classification Algorithms For Surface Defects Of Steel Plates And Strips

Surface defects is an important factor affecting the quality of steel plates and strips.According to statistics,more than 60%of the quality issues of domestic steel products are caused by surface defects.Automatic surface defect detection systems can perform online detection and timely feedback of surface defects,which is of great significance for improving the surface quality of steel plates and strips.With the improvement of production line speed and the increasingly strict requirements of users on product quality,it is urgent to improve the efficiency of defect detection and recognition algorithms,as a result of improving the detection speed and detection accuracy of surface inspection systems.It is an important research direction to develop a fast algorithm for detecting and identifying defects in different production lines and product surface conditions to meet real-time detection requirements in the field of surface inspection.In this dissertation,the algorithms of image segmentation,target detection and defect classification are deeply studied.Algorithms such as fast image segmentation,improved extreme learning machine and end-to-end obj ect detection are developed,which are applied to three typical steel products,including pickled steel strips,plates and hot rolled strips respectively.At the same time,focusing on sample collection problem of new production line and the difficulty of detecting unknown defects,the migration of defect samples of different production lines and the automatic identification method of unknown defects are studied.The main research contents and results are as follows:(1)Aiming at the complex surface background and variable defect morphology of the plate,a fast classification algorithm based on the combination of extreme learning machine(ELM)and genetic algorithm(GA)is proposed.Through the two evolution strategies of gene locking and dynamic mutation rate,the genetic algorithm accelerates convergence in ELM parameter optimization,and the final classification result is more stable.Experiments show that the recognition rate of common surface defects of 9 types of plates is 94.30%,which is 5%higher than the origin ELM algorithm.The algorithm can also meet the online detection requirements of plates with a low production line speed.(2)Aiming at the problems of fast transferring speed of hot rolling,variable surface morphology,scale interference,an end-to-end deep learning object detection method was introduced.In the training process of the model,the transfer learning method is adopted,and the feature extraction network trained under the large-scale data set can greatly improve the training efficiency.The algorithm transforms the traditional detection process of extracting the pre-selected areas and classification into a regression problem that predicts the position and category of the defect at the same time,which realizes the end-to-end object detection.Since the network only performs one calculation on the image,the calculation speed is fast.Experiments were carried out on 7 types of common hot rolled strip surface defects,with a mAP of 92.54%and detection speed reaching 14 FPS,meeting the requirements for online inspection of hot rolled strips.(3)Aiming at the characteristics of surface and defect shape of pickled steel strips and fast running speed of the poduction lines,a fast surface defect detection algorithm based on image segmentation with low level information is proposed.The algorithm accelerates the computational efficiency by introducing the integral graph,and applies the Hough transform method to the edge defect detection.The algorithm was applied to the surface defect detection of pickled steel strips.The detection accuracy of the surface and edge defects of the strip reached 97.9%and 95.2%,respectively,and the detection speed reached 50 FPS,which satisfies the requirements of online detection of pickling lines.(4)Due to the great differences in background and texture information of product surface images on different production lines,defects samples obtained on other production lines cannot be directly applied to the learning of new production line systems.Therefore,newly installed surface inspection systemsoften lack in sufficient defects samples.Collecting defect samples from the new production lines usually need a long time,which affects the detecting system usage.By using the generative adversarial network,the defect samples obtained on other production lines with the defect-free image of the new production line are combined to generate defect samples that can be used for the newly installed production lines,which achieves the purpose of quickly collecting defect samples.In order to quantitatively evaluate the migration effect of defect samples,a classification residual network was designed to judge the realness of the generated defect samples,and evaluate the original data set and the generated data set separately.The model achieved 96.0%and 95.87%defect recognition accuracy on the original dataset and the newly generated dataset,respectively,indicating that the newly generated defect samples have good learning effects.(5)Existing surface detecting systems generally can not recognize unknown defects and usually identify unknown defects as other similar known defects.a two-level defect classification modal was proposed to identify unknown defects.There are five categories in the first level according to the defects’ form,and each category is subdivided into several specific defect categories in the second level.For known defects,the defect categories in the second level are the defect types,while for unknown defects,the defect categories in the second level are blank.The modal was tested with seven known defect types and one unknown defect type of of hot rolled strips.The experimental results show that the false recognition rate of the new recognition algorithm for unknown defects is reduced from 61.3%to 8.03%,which greatly improves the recognition accuracy of unknown defects.