Research On Recognition And Evaluation Of Silicon Steel Stripe Defects Based On Transfer Learning

With the rapid development of China's economy,the demand for electric power resources has soared.Therefore,the construction of electric power infrastructure is particularly important.Silicon steel sheet is an important production material for cores in many power equipment,and its quality has a profound influence on the entire power equipment production chain.In the industrial production process,due to factors such as the production environment and the manufacturing process,the surface of the steel strip is liable to produce some defects.Therefore,it is very necessary to accurately identify the surface defects of the steel strip in order to improve the quality of the product so as to ensure excellent performance of the power equipment.In view of the low efficiency,low recognition rate,and complex feature engineering used in manual detection and traditional machine learning methods,this paper proposes a transfer learning method based on convolutional neural networks to achieve feature extraction automatically based on actual production requirements.And the method overcomes the difficulty of training small samples and obtains defect recognition accuracy far beyond the traditional method,up to 96%.The main work content is as follows:Firstly,image preprocessing is performed to build a standard image data set.Above all,the problem of light reflection encountered in image data acquisition was introduced,and a solution similar to the integrating sphere device was proposed.Due to the constraints of the production line environment,industrial cameras cannot be mounted on vertical strips,resulting in distortion of the captured strip images.This paper adopts the perspective projection transformation method to complete the geometric correction of the images.This paper adopts the perspective projection transformation method to complete the geometric correction of the images;then the Canny edge detection algorithm is used to trace the edges of the image defects,so compete edge information is collected;Then,based on the identification of defect profiles,the area seed filling algorithm is used to obtain the area of various types of defects,which helps manufacturers to determine the quality of the steel strips according to the ratio of the area occupied by the defects of the steel strip.Then manufacturers can set different prices to maximize profits.Finally,based on the defect area and experience judgment,the defect image is tagged and a standard data set is constructed.Secondly,we studied the convolutional neural network algorithm and transfer learning theory.The principles of convolutional neural network model construction,parameter updating,model training methods and specific application details in the image recognition problem are described.Then the end-to-end model based on convolutional neural network is constructed.The model implements the automatic operation of feature extraction and can gradually abstract the underlying features of the image into high-level semantic features of the image,thus clearly demonstrating the various patterns of defects.In addition,the transfer training method based on the pre-training model effectively overcomes the difficulties of small sample training,prevents overfitting of the model,and greatly improves the accuracy of defect recognition.By comparing the difference between this method and the classic convolutional neural network algorithm,we can find that this method is better.Not only the recognition rate is higher,but also the operation is faster and the time cost can be effectively reduced.Finally,fine-tune the model structure to achieve higher defect recognition accuracy.Trimming the parameters of the pre-training model is used to realize fine-tuning of the model,and then the characteristic expression ability of different convolution layers and the influence on the performance of the model are observed,so that a targeted layer-based freezing of the image data of the silicon steel strip defect is adopted.The technology reduces the training time of the model and enhances the robustness of the convolutional neural network,which further improves the recognition accuracy of the model.Ultimately,the model can achieve accurate prediction of defect categories and the ability to obtain defect areas.