Scene Classification Based On Deep Convolutional Neural Networks

Scene classification refers to identifying the scene automatically according to the content contained in the image.It is one of the basic tasks in the computer vision field and the basis of several computer vision tasks,such as target detection,target tracking,image retrieval and so on.With the rapid development of digital multimedia technology and the wide popularity of computers and mobile phones,a large number of images are generated every day.There is an urgent need for the efficacious management of images.Therefore,the scene image classification is an important task both in the fields of academia and industry.In recent years,with the rapid development of the deep learning,convolutional neural networks have made a breakthrough in the scene classification task.The paper studies the multi-scale convolutional neural network and proposes two scene classification methods based on deep convolutional neural networks.The main contributions are as follows:Firstly,owing to the existing scene classification algorithms still have many shortcomings,an end-to-end multi-scale convolutional neural network is proposed for the scene classification.In the existing scene classification models based on convolutional neural networks,the multi-stage strategy is usually adopted.The convolutional neural network is used as the feature extractor to extract the high-level semantic features of the scene images from multiple scales,and then the traditional feature encoding method is utilized to generate the multi-scale features.It makes the feature learning and classifier training isolated and contraries to the end-to-end idea of deep learning.Based on the shortcomings of the existing algorithms,an end-to-end multi-scale convolutional neural network is proposed.The multi-scale scene patches extraction,feature learning,feature fusion and classifier training are unified into the proposed algorithm.Additionally,the algorithm employs a multi-tasking strategy to fully explore the performance of the scene features at each scale.In the experimental parts,the proposed scene classification method based on end-to-end multi-scale deep convolutional neural network achieves the accuracy of 80.9% and 94.5% respectively in MIT 67 indoor scene database and Scene 15 database.Secondly,according to the complexity of the content contained in the scene image,the paper proposes a scene classification method based on discriminate scene features.The discriminative patches of scene images are extracted from the original scene image by an iterative algorithm.The scene patches can effectively represent thecurrent scene category,and obviously distinguish from other scene categories.The proposed model takes both local and global features of the scene into account.On the one hand,the global feature of the scene image is extracted by the convolutional network.On the other hand,the local feature is extracted from the scene image patches by the improved locally aggregated descriptor.Finally,combining the global and local features of the scene,the proposed method achieves the accuracy of 83.2%and 94.8% respectively in MIT 67 indoor scene database and Scene 15 database.