Scene Classification Based On The Deep Learning

With the development of multimedia technology and popularization of network, people have access to an increasing number of digital images in ordinary living. In the face of huge image data, the traditional images management by manual annotation is no longer feasible. Architectures are applied in many existing machine learning algorithms including support vector machine, neural networks with only one hidden layer, kernel regression and many others shallow architectures. Those shallow architectures with limited samples and finite computing units are incapable of representing the complex function and place restriction on the generalization capability of classifying complicated issues.In recent years, the methods of Deep Learning with the neural network have achieved many breakthroughs, and get a large number of innovative applications in computer vision, speech recognition, natural language processing and other fields. The Deep Learning is a process of simulating human brain learning mechanisms. It uses multilayer neural network to express abstractly the real object or other data of voice or text. This method combines the features of the extraction and classifier to a learning framework, classifying related objects recognition, etc. This deep nonlinear structure of the network can be the approximation of complex function, distribute according to the input data, extract the essence of features from the input data.To avoid traditional requirements of manual design features and improve the robustness of features, this paper leverages deep learning for scene classification and raises a new deep convolutional neural network model. It can make full use of the advantage of deep convolutional neural network and extract a variety of scene features form the databases. According to the characteristics of the deep convolutional neural network with hierarchical information extraction, this model uses smaller convolution kernels to extract more of the low-level image features and laid a good foundation for high-level features. At the same time, the scene models of been trained with last layer depth convolution neural network of 4096 neurons, combining with the Lib-SVM classifier to classify the original scene image. Through the experiments on two datasets show that the method of deep convolutional neural network can extract the image features effectively, and the model of scene has stronger generalization performance and classification efficiency highly.