Scene Classification Based On The Convolutional Neural Networks

With the rapid development of multimedia and Internet technology,the scale of digital images has expanded dramatically.Facing the massive complex scene image resources,how to classify and manage them effectively,to facilitate the incremental storage management of image resources and the accurate retrieval of users,has become an urgent problem to be solved in the field of machine learning and image retrieval,scene classification technology came into being in this case.In order to avoid manual design features in traditional scene classification technology and improve the robustness of features,this paper take advantage of deep convolutional neural network to extract features for classification based on deep learning technology,and mainly studies the following aspects.1)Transfer learning.In this paper,we use the deep convolutional neural network — CaffeNet model,which is trained on the large scale dataset ImageNet,to extract the preliminary scene image feature,and then the final feature is obtained by the method of principal component analysis,finally the final feature is fed into an support vector machine classifier to predict the class lable.Amoge them,the characteristics of different layers' features in the network and the influence of dimension reduction by principal component analysis on the classification accuracy are studied.2)Fine-tune CaffeNet model.For small scale scene dataset,due to over fitting problems,it is difficult to train a deep convolutional neural network from scratch.In this paper,we fine-tune the last fully connected layer,then the parameters of pre-trained CaffeNet are used to initialize fine-tuned model,finally,according to the characteristics of different layers,different learning rates are set up to train the whole model,in this way we can obtain a more domain-specific model.3)Feature fusion.The deeper the network level of convolutional neural network is,the more it can capture the global features of the image.However,it will inevitably lose the local features hidden in the lower layer of the network.In this paper,aiming at the shortcomings of only using single feature of high layer,we reduce convolution layers' features dimension by principal component analysis,then all layers' features after L2 norm normalization are fused,so as to further improve the robustness of feature.