Research On The Classification Of Diamond Saw Blade Crack Recognition Based On Depth Learning

Diamond saw blade in the production process due to the nature of the material structure,heat treatment unreasonable,heating / cooling and other factors lead to surface cracks.Seriously affect the product quality and service life,the use of the process may cause serious harm to human body,so the diamond saw blade surface crack identification very practical significance.The traditional method based on machine vision is to define the features by experts,design the corresponding feature extraction and classification algorithms to identify,and the definition of features is influenced by experience.When the artificial design algorithm can not express the advanced features of images,the recognition rate will be greatly low.Deep Learning Deep neural networks are used to automatically learn deep-seated features from big data samples,instead of manually constructed features,and to describe features more appropriately so as to improve the recognition rate.In this paper,based on the deep learning convolution neural network to realize the identification of diamond saw blade crack classification.Aiming at the problem that the noise in the image interfered with the learning features of CNN model,the recognition rate was low and the training speed was drastically slowed down.Pretreatment stage to study the denoising algorithm improvements.For a single denoising algorithm can not effectively remove hybrid noise,and image information loss serious problem.In this paper,an improved bilateral filtering algorithm and an NLM filtering algorithm are proposed.The former can remove mixed noise and the latter can preserve the structure information.On the basis of this,a second improved bilateral filtering algorithm is proposed.After the algorithm preprocessing the image,the model training time is greatly reduced,and the recognition accuracy is greatly improved.When the CNN model is used to identify saw blade cracks,there is a problem of over-fitting the network and slow convergence.This article makes optimization and improvement.For the over-fitting problem,the sample is enhanced and expanded,and the sample set is expanded to improve the generalization ability.The algorithms of algorithm optimization and gradient descent are used to update the hidden layer and Softmax classification layer parameters to increase the regularization term.The Dropout strategy is adopted to study the learning rate Adaptive update.As for the convergence speed,the activation function Sigmoid and Re LU are fused.The improved activation function preserves the rapidity of Re LU and the smoothness of Sigmoid.After the optimization and improvement,the accuracy of crack identification is improved..