| Sheet and strip products are widely used in the fields of construction,transportation and engineering machinery.With the continuous improvement of market requirements,the surface quality of plate and strip has become an increasingly concerned issue to steel companies and users.Traditional eddy current testing is susceptible to high frequency signal interference.Magnetic flux leakage detection and laser detection are not capable of detecting small defects.Machine vision inspection methods based on image processing are gradually widely used due to their advantages such as accuracy,efficiency and low cost.Convolutional neural network(CNN)is an algorithm that can autonomously mine image features and learn,and can been widely used in the field of image processing.However,the conventional CNN uses a random convolution kernel method.It cannot control the direction of learning features.It can only control the fitting process by setting hyperparameters.When adjusting the optimal hyperparameters,the model will fall into the learning bottleneck if the accuracy cannot be further improved.In order to solve the above-mentioned limitations,machine vision technology is used to study the algorithm for identifying surface defects of steel plates in this paper.An improved algorithm based on feature fusion CNN is proposed to further improve the accuracy of surface defects identification on steel plates.The main contents are as follows:(1)The theoretical basis of convolutional neural networks was explained.The mechanism of convolution operation and the advantages of CNN in processing image data were described.The basic structure and training process of CNN were analyzed,which provides theoretical support for the establishment of defect identification models.(2)The defect identification model was established and optimized.A database of 1200 samples of steel surface defects was collected and labeled,and the processing effects of different filtering methods were compared through experiments to determine the use of Gaussian filtering to eliminate noise.The data enhancement method was used to simulate the actual distribution of defects,and the database capacity was expanded.By comparing the fitting results of the defect data using different network models,the VGG16 structure was selected as the main network structure of the defect detection model.The reason for overfitting phenomenon of the model was discussed,and the Dropout regularization scheme was selected to avoid over-fitting.The influence of different hyperparameters on data fitting was analyzed,and the best combination of hyperparameters was determined through experiments.The identification accuracy of CNN can reach 92.55%.(3)Aiming at the bottleneck of network learning,a multi-channel feature fusion algorithm was proposed.By the experimental analysis of different feature operators on the image and the exploring of the influence of the fusion method of different combined feature operators on the model,the best fusion scheme was obtained.In order to reduce the calculation parameters of the network,different ways of single-channel conversion weight ratios were explored,and the best conversion weight was obtained through experimental comparison.Using the method of feature activation map and heat map analysis to visualize the model learning effect before and after feature fusion processing,the effectiveness of the feature fusion method was verified.The recognition accuracy of the feature fusion CNN can reach 95.63%.(4)The steel sheet surface defect detection system was developed and applied to a domestic plate and strip production line.The hardware of the steel plate surface defect detection system was designed based on the light conditions at the scene.The application software of the steel plate surface defect detection system was developed,including the user operation interface and the functions such as automatic acquisition,image splicing,and defect suspicion area detection,achieving good application results. |
PDF Full Text Request