Image Classification Algorithms Based On Two-stream Convolutional Unit

As the basis of artificial intelligence,image classification technology plays an extremely important role.With the application of intelligent devices,the demand of image classification technology is higher and higher.Most of the traditional image classification algorithms rely on hand-crafted features and classifiers,which have some limitations,and are affected by illumination,background,scale,variable viewpoint and some other factors.In recent years,with the rapid development of deep convolutional neural networks,the image classification technology based on convolutional neural networks has made a breakthrough.Its performance of classification surpasses the performance of traditional image classification algorithms.It is widely used in industry and attracts more and more researchers.Now the convolutional neural network is getting deeper and wider,but this may bring a large number of parameters and increase the training difficulty.Although the neural networks based on simplified convolution can reduce the network parameters,some information is discarded which decreases the performance of the networks.In order to solve this problem,a two-stream convolutional unit is proposed.The convolutional unit contains two different types of filters,one is for extracting the features that contain the information in the channels and the other is for extracting the features that contain the information across the channels.In order to verify the two-stream convolutional unit,a deep convolutional neural network called CTsNet(Cascaded Two-stream Network)is constructed by replacing the traditional filter with the convolutional unit and combining with the operation of pooling and activation.Experimental results on CIFAR database show that the proposed two-stream convolutional unit can extract features containing the information in and across the channels separately.The diversity in features can be increased and more information can be preserved.At the same time,the network can improve the network performance with fewer parameters.The traditional ReLU(Rectified Linear Units)turns the negative numbers to zero and remains the positive numbers which can accelerate the convergence speed of network training.In order to make full use of the positive part,we propose a self-ReLU,which is multiplied by a certain proportion between zero and one so that different activation values have different effects.By using this activation function in the NIN network structure,this algorithm is validated on the CIFAR database.The experimental results show the effectiveness of this algorithm.