Research On Image Classification Algorithms Based On Convolutional Neural Network

Nowadays,with the rapid development of Internet and multimedia technology,a large amount of image data are generated.Effective classification of these data can not only improve the efficiency of information querying,but also prepare for other computer vision research such as target detection and image retrieval.In recent years,Deep learning technology has made great progress in the field of computer vision,especially in image classification.By simulating the human brain's visual nervous system,the original image is abstracted to get high-level semantic features that's more close to the objects,so as to achieve good classification.In this paper,image classification is studied based on Deeping learning.The main contents are as follows:(1)Studied the current situation of image classification at domestic and foreign,including traditional machine learning algorithms and current mainstream image classification algorithms based on deep learning.The advantages and disadvantages of these algorithms are analyzed.At the same time,the related theoretical knowledge of convolutional neural networks is elaborated,with emphasis on its basic structure and principle.(2)To solve the problem of expression ability and recognition effect of convolutional neural network affected by activation function,an improved activation function T-ReLU is proposed and applied to image classification of convolutional neural network.The negative half-axis of the function inherits the left half of Tanh function and introduces adjustable parameters ? to make its left side have soft saturation characteristics.The positive half-axis uses the linear part of ReLU function to avoid gradient disappearance during model training.The experimental results show that T-ReLU functions can effectively improve the classification performance of convolutional neural networks compared with common activation functions such as ReLU.(3)Aiming at the problem that image classification algorithm based on deep learning can't effectively fuse multi-level features and the classification accuracy of model is poor,an image classification algorithm based on multi-level depth features is proposed.Firstly,an 11-Layer convolution neural network with residual units is designed and constructed.At the same time,a batch normalization algorithm is introduced to normalize the image data after each convolution operation.Subsequently,the output features of the last pooling layer in the convolution neural network are effectively fused with that of the connecting layer based on the cascade idea to make the network extracted features more diverse and discriminant.Finally,the fused features are classified by Softmax classifier.Compared with the current classical image classification algorithms(such as NIN,DSN,Highway Network,etc.),the experimental results show that our algorithm has higher classification accuracy.