Image Recognition Based On Convolutional Neural Networks

The biological neural network is made up of massive neurons connected by dendrites and axons,which can be continuously studied and evolved.The artificial neural network is a mathematical model which is based on the basic principle of neural network in biology and the knowledge of network topology.It simulates the mechanism of human brain's neural system to external information processing mechanism.Artificial neural network research has a period of rapid development at the end of the last century.Due to various conditions,it fell into a trough after that.Deep learning is based on artificial neural networks.It is a deeper and more complex network model.In recent years,due to the rapid development of computer technology,especially the improvement of hardware conditions,deep neural network's training time is greatly reduced.Deep learning has quickly become a research hotspot.In the field of digital recognition,speech recognition,self-driving car,image recognition and other fields,the deep learning technology has a wide range of applications.A new generation of artificial intelligence technology,represented by deep learning,is gradually penetrating into people's lives and promoting social development.Convolutional neural network as a typical of the deep learning technology,in recent years has also made rapid development.Convolutional neural network is a special deep neural network,which contains a special convolutional layer and down-sample layer.The convolutional layer's neurons use local connections and weight sharing way to connect with previous layer,thereby greatly reducing The number of parameters to be trained.The down-sample layer can greatly reduce the input dimension and reduce the network complexity,making the whole network higher robust,and can effectively suppress the over-fitting problem.These designs make convolutional neural networks have unparalleled superiority in identifying tasks such as scaling,displacement,and other forms of distorted 2D graphics.In 2012,the Alex-net get great success in that year's image-net competition.And in 2015image-net competition,the residual neural network firstly appeared.The residual neural network is also a new deformation of the convolutional neural network.It doesn't learn directly about the target,but instead learn a residual.Because of the appearance of residual neural network,convolutional network research has entered a new stage.In this paper,an improved multi-scale residual neural network is proposed based on the traditional residual neural network.Our network architecture allows the convolutional layer to "observe" data from multiple different scales to achieve richer input characteristics.And the depth of the network is reduced,which effectively suppresses the occurrence of gradient disappearance and reduces the difficulty of training.Due to increasing the number of training parameters,we make the network learning ability becomes stronger.We adjust the zoom parameter,the depth of the network,the number of learning modules in each group,the position and value of dropout.In the end,the accuracy of our network structure is higher than that of thetraditional residual neural network.Finally,we use the ensemble learning method-the majority voting to reduce the classification error rate of the network.We get 3.49%error rate on the CIFAR-10 dataset,which is about 3% lower than the original residual neural network.