Scene Classification Based On Deep Learning And Spatial Visual Bag-of-Words Model

With the development of information technology,scene recognition and classification have a wide range of application prospects and theoretical significance in the fields of image etrieval,intelligent navigation,video surveillance,and other computer vision tasks,making a higher demand for more effective scene classification technology.However,it is relatively difficult to process and identify the scene image due to the variousness,fuzziness and various lighting and scale conditions of scene images.Therefore,how to make full use of the information contained in the image to carry out independent and rapid learning so as to identify the scene image quickly and effectively,has become an important topic in the field of computer vision.Therefore,in order to reduce the limitations of traditional classification technology,traditional bag-of-visual-words model and deep learning method are used in this thesis,so as to develop the powerful learning ability and learn more hierarchical visual information from the scene images.Three main research aspects as below are contained in this thesis:First,in the case of insufficient performance of traditional underlying features,the CNN convolution layer has a powerful learning ability of the image features,and replaces the traditional local features to obtain more information of the image.In this thesis,a fast multi-level convolution neural network model is designed and trained,and the features of different levels are compared and analyzed to reconstruct the characteristics of the dictionary.Second,for the lack of spatial information of the traditional bag-of-visual-words model,combined with the idea of the spatial pyramid model,the extracted convolution feature is coded by the improved soft-assignment coding method to generate the space visual dictionary.And training different Kernel classifiers to achieve better accuracy.Third,a scene classification system for the experimental and algorithm testing is designed and implemented in this thesis,including the basic input and output functions,algorithm selection,the test of image classification,the display of results and other functions.By the comparative analysis of the experiment,it is shown that constructed network model has realized the multi-level data extraction and improved the expression ability of the features.It combined with the spatial visual lexicon model to make up for the lack of spatial information,which providing a new approach for improving the accuracy of the scene classification.