Research On Classification Of Architectural Style Image Based On Convolution Neural Network

Deep learning is a new field in machine learning research.At present,the models of deep learning network mainly have deep belief network,limited boltzmann machine,stacked autoencoder and convolution neural network.In recent years,convolution neural networks have achieved incredible achievements in different fields,including image classification,target and face detection,speech recognition,and so on.In this dissertation,the convolution neural network is the most important factor in image recognition.One of the most notable successes is their performance on ImageNet dataset(Krizhevsky et al.,2012),where they have no proper competitors and win second year in a row.This dissertation mainly focuses on the network design and parameter optimization of convolution neural network.The effect of convolution neural network on image classification is mainly determined by the number of network layers and network parameters.How to design the number of convolution layers,the number of hidden layers and good parameters in the convolution neural network are key points in determining whether the convolution neural network model is successful.Therefore,the main content of this dissertation includes the following works:(1)First of all,this thesis designs a shallow convolution neural network model of 5-layer convolutions,tests them in ten and twenty-five architectural style datasets,and uses the relu function instead of the traditional sigmoid function as the activation function and trains the model using Dropout to optimize the network parameters,thereby enhancing the classification of building style images.(2)Given the complexity of architectural style image classification.This is followed by the design of a 9-layer network layer of deep convolution neural network structure,and then implements the test on the ten and twenty-five architectural style datasets.Through the use of maxout function as an activation function,and the use of DropConnect in training process to optimize the network parameters,so as to improve the classification of building style images.(3)The two kinds of network models proposed in this thesis are experimentally studied.The experimental results show that the deeper the network layer of convolution neural network,the more comprehensive the feature of the image,the better the training classification effect.However,the more the network layer,the more network parameters,the longer the training time,the smaller the improvement of the accuracy of classification,given many factors,it is a proper choice to set the network layer to 13.This thesis respectively does experimental study on two kinds of convolution neural network classification models put forward,and analyzes the experimental results and summarizes some practical rules of depth classification on image classification,which have a certain reference to solve practical problems.