Research And Application Of On-line Detection Algorithm For Surface Defects Of Cold Rolled Strip Steel

Surface defect is an important and visible factor affecting strip quality.According to statistics,most strip product quality complaints are caused by surface defects.The on-line inspection system of surface defects can detect and feedback surface defects online,which is of great significance for improving the surface quality of strip steel.With the continuous improvement of production line speed and the increasingly strict requirements of product quality,the detection speed and accuracy of surface defect on-line detection system also need to be improved.The online surface defect detection system includes two parts:hardware integration and software flow.The core of the software flow is the detection algorithm Therefore,the core of improving the speed and accuracy of the detection system is to improve the speed and accuracy of detection algorithm.In this paper,the defect detection algorithms such as defect area segmentation,defect feature extraction and defect classification are studied in depth.Aiming at the characteristics of fast running speed of cold rolled strip steel production line,the fast defect segmentation algorithm of four-neighborhood difference,Block-LBP fast defect feature extraction algorithm and Threshold-ELM fast defect classification algorithm were developed respectively.These algorithms are integrated into the on-line detection system.The main research contents and achievements are as follows:(1)To solve the problem of complex logic and slow speed of common segmentation algorithms,a fast four-neighborhood difference segmentation algorithm was proposed.This algorithm borrows from the SURF algorithm,which uses integral image blocks to represent local area features.Firstly,the defect images are divided into sub-blocks,and the mean grayscale of sub-block area is used to represent the regional feature of the sub-block.Then the differences between four-neighborhood sub-blocks are obtained to represent the difference between blocks.Finally,by setting the appropriate high and low thresholds and appropriate the range of area growth,a more accurate defect area is finally obtained.The experimental results show that the detection rate of defects of the four-neighborhood difference algorithm is 97.4%,which meets the accuracy requirements of on-line defect detection.The average running time of the algorithm is only 5 ms/amplitude,which has great advantages over the traditional SIFT and SURF algorithms.It not only meets the online detection speed requirements,but also leaves enough time for the subsequent feature extraction and classification steps.(2)Aiming at the problem that LBP and its various improved modes can only sample a single pixel,unable to extract a large range of texture features,and susceptible to noise and irrelevant detail features,a fast Block-LBP feature extraction algorithm is proposed on the basis of LBP.The Block-LBP replaces the grayscale of a single pixel with the average grayscale of a sub-block containing multiple pixels,so as to suppress the noise and extract texture features in a larger range.The experimental results show that:compared with SIFT,SURF and other common feature extraction algorithms,the Block-LBP has obvious advantages in running time,reaching 10.9 ms/amplitude,and meeting the speed requirements of online defect detection;at the same time,the accuracy of defect classification has been greatly improved,reaching 95.94%,which meets the accuracy requirements of online detection.(3)Aiming at the problem that ELM algorithm cannot recognize the new unknown type of defects in actual production,this paper introduces the threshold strategy of maximum classification value and proposes an improved Threshold-ELM algorithm on the basis of ELM algorithm.The threshold strategy tests the classification result with the set threshold value.Only when the maximum classification value in the classification result vector is greater than the threshold,the sample should be classified to the corresponding class;otherwise,it is determined that the sample belongs to an unknown type.The severity of detecting unknown types of defects is changed by changing threshold.The experimental results show that Threshold-ELM has obvious advantages in both classification time and training time compared with BP,SVM and other common classification algorithms.When new unknown type of defects appear,threshold-ELM can check out 93.33%new defects,which effectively alert the generation of unknown defects.After using the retrained classification model,the classification accuracy of threshold-ELM reaches 92.67%,which effectively classifies new types of defects.(4)In order to apply all above rapid detection algorithms in actual production environment,this paper design of a new inspection system based on the selection of hardware and software.The core algorithm of the system integrates the four-neighborhood difference defect segmentation algorithms,the Block-LBP feature extraction algorithm,and Threshold-ELM defect classification algorithm.The system is experimented on the complete large cold rolled strip steel samples.In order to realize the online detection in the actual production environment,the four-neighborhood difference defect segmentation algorithms,the Block-LBP feature extraction algorithm,and Threshold-ELM defect classification algorithm proposed in this paper were integrated into the online detection system of surface defects proposed by our research group and then applied to the actual production line.Experimental results showed that the average accuracy of the detection system is 95.38%,average running time is 22.56 ms/amplitude,94.11%of new unknown defects can be checked out,all results have reached on-line inspection standard,which prove that the system can meet the requirements of on-line defect detection in the actual production environment.