Research On Surface Defect Detection Method Of Steel Plate Based On Image Processing Technology

The steel industry is one of China’s pillar industries,with its export volume ranking among the top in the world.But with the further improvement of steel quality requirements,especially steel plate as a very common commodity used in the steel industry,its surface defects have always existed,which will cause many product quality hazards.Therefore,the surface defect detection technology of steel plate is of great significance to the steel industry and the whole society.In view of the low accuracy and slow detection efficiency of traditional steel plate surface defect detection methods,this article uses image processing technology to study and analyze the steel plate surface defect detection methods,preprocess the steel plate image,and introduces the improved crow optimization algorithm to segment the defect image,realizing the identification of steel plate surface defects.To solve the steel plate surface defects difficult to accurately identify the problem.The research content of this article is as follows:(1)Through analyzing the iron and steel production process as well as its defect characteristics,a defect detection system design scheme has been established.This includes selecting an industrial camera,lens and light source,constructing the hardware platform,and designing the software process.To mitigate the impact of lens-induced image distortion on detection accuracy,a camera calibration experiment and image correction were conducted.Furthermore,to address the issue of neural networks struggling to capture valuable defect features due to an insufficient number of defect samples,a dataset of steel plate surface defect images was created.(2)To enhance the accuracy of steel plate surface defect identification,a study was conducted on pretreating the stecl plate defect images.This involved converting the color image into grayscale to eliminate extraneous data and reduce algorithmic processing time.To solve the problem that the image contains too much noise affecting the subsequent processing,an improved adaptive median filtering algorithm is proposed.By performing a quantitative analysis of the experimental results,a comparison was made to highlight the benefits of the improved adaptive median filtering method.To alleviate the issue of suboptimal image acquisition conditions and uneven illumination,three image enhancement techniques(i.e.,histogram equalization,guided filtering,and multi-scale Retinex algorithm with color restoration)were evaluated and analyzed.A performance experiment concluded that the multi-scale Retinex algorithm with color restoration was the most effective approach for enhancing steel plate surface defect images.(3)To address issues associated with conventional image segmentation methods(e.g.,substandard segmentation outcomes,reduced efficiency,and increased susceptibility to external factors),this study details a two-dimensional(2D)Otsu image segmentation tech n ique that incorporates an enhanced Crow algorithm.A multi-strategy optimized Crow mcthod is presented as a solution to the challenges surrounding low convergence performance and local optimization.Numerical simulation analysis of test function shows that the improved Crow algorithm has higher convergence speed and solving accuracy.The two-dimensional Otsu image segmentation method of the improved Crow algorithm is used to segment the preprocessed steel plate defect images.The experimental findings demonstrate that the algorithm presented exhibits superior performance in terms of both segmentation speed and accuracy.It lays a foundation for the accurate classification of steel plate defects.(4)This article proposes a steel plate surface defect recognition method that employs a backpropagation(BP)neural network model to address inaccuracies associated with conventional approaches.Through the feature extraction of steel plate defect area,a total of 25 features were obtained,including gray level,geometric shape,texture and projection.To reduce data dimensionality,this study applies the Fisher criterion method to identify the 15 features with the highest Fisher ratio.Subsequently,a classification and prediction model is developed using BP neural network techniques.Ultimately,the classification model achieves a recognition rate of 85.69%when applied to test samples.This outcome demonstrates that the model is effective under basic feasibility conditions and confirms the efficacy of the selected features.