Research On Steel Surface Defect Recognition Method Based On Deep Learning

With the development of national economy,steel become the main part of the high efficient energy-saving construction steel.steel surface quality affects the use of safe and sales,quality testing on the surface of the steel have been an important process in production.Deep learning is the essence of artificial intelligence algorithm and is widely used in many fields.In this paper,the deep convolutional neural network(CNN)is used for the quality detection of the surface of the steel to improve the detection level.In order to train a more robust model with better learning ability,firstly makes data enhancement on the collected image data set of the shape steel.A large number of data is very important for training neural network model,which can increase the nonlinear expression ability of the model and reduce the condition of over-fitting.In this paper,the data set of shape steel can be expanded artificially in combination with the possible situation in the process of section steel production.Then establishes a linear CNN to classify the surface data sets of sections.The linear CNN has simple structure,convenient network level and convenient parameter adjustment.Build different depth of the CNN to verify the effect of different network structure,comparing different influence on model,data volume and adjust the parameter to achieve good recognition effect.In the end,a new structure of CNN is designed to classify the surface data sets.The new CNN structure is not only on the vertical depth increase,also on the horizontal width of the increase,use 1 * 1,1 * 3 filter,BN layer to reduce the difficulty of training,verify the network structure of different complexity and the recognition accuracy of different size datasets..And adjust the parameters to improve the model recognition effect.