Research And Application On Convolutional Neural Network Algorithm Based On Improved Activation Function

The research and application of big data and deep learning have received increasing attention in various fields with the development of machine learning,artificial intelligence,and network technology.Convolutional neural networks occupy an extremely important position in deep learning research because of its unique characteristics of local connection and weight sharing,and because the convolution process in convolutional neural networks can automatically and quickly learn the characteristics of images,therefore,it shows great advantages in image classification tasks.With the in-deep research on deep learning and image classification,convolutional neural networks still have some shortcomings in image classification tasks.For example,in different tasks,the same model classification ability is different,model structure is not uniform and universal;in the stage of establishing the classification model,the convolutional neural network used to implement image classification tasks does not have a universal model structure with strict theoretical support,in order to determine the structure of the classification model,it is necessary to select and optimize through a large number of experimental studies,the convolutional neural network requires strict scale of training samples,once the number of training samples is small,the classification accuracy of the model will be very low.Since the activation function plays an extremely important role in the convolutional neural network algorithm,if the selection is not correct,it will reduce the classification accuracy of the model.Whether the selection of the activation function is appropriate or not directly affects the performance of the network classification.In the convolutional neural network,different activation functions have their own advantages and disadvantages.For example,the linear activation function can not make the model have good multi-classification function,and the nonlinear saturated S-type activation function can meet the needs of multi-classification,but it is easy to cause the problem of gradient disappearing in network training,the nonlinear piecewise function Relu can avoid the problem of gradient disappearance,but it is easy to be the problem of “neuron die” in the network,etc.How to choose and use the best activation function is a big difficulty in image classification based on convolutional neural networks.Aiming at the problems existing in image classification of convolutional neural networks and the shortcomings of common activation functions,this paper studies and improves the activation function of convolutional neural network algorithm and network model structure,in order to improve the accuracy of image classification and increase the stability and versatility of the model.The specific research is as follows:(1)Through the study of the activation function,the advantages and disadvantages of various commonly used activation functions in the neural network are analyzed,the strong suppression of the negative excitation of the Relu function on the network,and the gradient disappear of the hyperbolic tangent function Tanh to the network,the activation function T-Relu function with soft saturation-linear structure is proposed,and an improved convolutional neural network algorithm TRCNN is proposed.According to the theoretical derivation of the activation function and algorithm,it is proved that the soft saturation-linear structure activation function T-Relu can make the TRCNN algorithm have better feature learning ability.,(2)Based on the knowledge and research of convolutional neural network model,through a large number of experiments,the model structure of convolutional neural network based on TRCNN algorithm is gradually determined.The basic structure is C-CS and the network depth is 8 layers,using the combination of dropout-L2 to prevent overfitting,and successfully constructing the image classification model Typ-CNN,which is applied to the classification of self-built typhoon-grade satellite cloud maps,the classification accuracy reaches 83.72%,Compared with the traditional CNN model,the classification accuracy is 2.3% higher,which proves that the Typ-CNN model has a good classification effect.(3)Finally,through the comparison of multiple experiments on the Typhoon dataset,MNIST dataset and CIFAR-10 dataset,the classification accuracy rates of 84.17%,98.64% and 72.34% were obtained respectively,which is better than other activations under the same conditions.The classification effect of CNN algorithm under the function.The availability of the TRCNN algorithm based on the T-Relu activation function is demonstrated.Based on the results of all experiments,the CNN algorithm based on different activation functions has different classification ability under the same data set and classification model structure.The classification model based on the same algorithm and network structure has different classification effects on different data sets.