Image Classification Algorithm Based On Improved Convolutional Neural Network

In recent years,with the rapidly development of deep learning.Image classification related algorithm research has gradually become a hot topic for researchers.It has a broad development space in the field of industrial product defect detection.Traditional defect detection methods have low recognition accuracy and weak generalization ability,which can not meet the needs of industrial production.Convolution neural network realizes self-learning and classification by feature extraction of input pictures,and achieves amazing results in many competitions.Researchers improve the performance of neural networks by increasing the depth and the number of feature planes.But the more complex the network model,the more difficult it is to train.In the actual industrial application,not only the accuracy of defect detection,but also the detection speed of network model should be considered according to the pipeline.In view of the above problems,this paper studies the defect detection algorithm based on the improved convolution neural network,and the main research contents are as follows:(1)In this paper,the development process of convolutional neural network is deeply understood,and then the layers of convolutional neural network and the relationship between them are described.According to the difference between the convolution neural network and the depth neural network,and according to the back propagation algorithm of the depth neural network,the back propagation algorithm of the convolution neural network is derived.The forward propagation algorithm and back propagation algorithm of convolutional neural network are summarized.(2)In view of the phenomenon that the weight probability of the current common pooling algorithm is easy to lead to the loss of features,this chapter proposes a rejection acceptance pooling algorithm.It combines the rejection sampling method to select from the polynomial distribution of the activation in the pool area,and adopts the random process to ensure that the maximum activation has the chance to be selected and transferred to the network,while ensuring that the strong activation has a higher chance to be sampled.Through the experiments of different pooling algorithms on different neural network models,it is proved that this pooling algorithm can greatly improve the accuracy of defect detection of the magnetic data set.The experimental results on the open data set cafir-10 show that the algorithm has certain universality.(3)In order to improve the performance of convolutional neural network by improving the activation function,this paper proposes an efficient convolutional neural network algorithm which can automatically select the best activation function.The algorithm builds Disk Net based on VGGNet-19 network model,which uses WP-model to predict and select the best activation function.In WP-Model,firstly,Markov chain Monte Carlo(MCMC)is used to deduce the prediction value and determine the prediction probability,then the effective information is extracted from VGGNet-19 training curve to determine the evaluation point(EP).In the process of Disk Net training,when the prediction probability is higher than the threshold value,the neural network will select the current activation function.If the number of training exceeds the EP point and does not reach the threshold value,the original activation function will be used.The experimental results show that the accuracy of the improved convolution neural network model is 96.9%.