Convolution Neural Network And Its Application In Feature Learning Of High Resolution Satellite Remote Sensing Images

In recent years,deep neural networks have achieved remarkable results in image classification and recognition,natural language processing and other fields with its powerful feature extraction and representation skill.Among them,convolutional neural network has characteristics such as local receptive fields and shared weights,which not only reduces the computational complexity of image feature extraction,but also effectively represents the overall features of an image by certain combination of local features.However,there are still many bottlenecks in applying convolutional neural network to remote sensing.The main problem is that high resolution satellite remote sensing images contain a large amount of information,and high image complexity.This poses a great obstacle to the feature extraction based on convolutional neural network.To solve this problem,we use a large number of remote sensing data,design a convolutional neural network structure suitable for high resolution remote sensing image feature extraction,and apply the extracted features to classification,detection and caption of remote sensing images.This paper firstly introduces the basic structure of neural network,followed with the learning algorithm of neural network with convolutional neural network as an example.Meanwhile recurrent neural network models are presented.After that,the feature extraction ability of different convolutional neural network is introduced.The gradient learning algorithm is applied to train convolutional neural network models on the MNIST handwritten digit database.We analyze the convolution kernel,the network depth and other factors that could affect the image feature extraction in network structure.Experimental results show that network structure with deeper network,smaller convolution kernel size,more convolution kernels and smaller pooling size has a better feature extraction ability and performance in classification results.This kind of network structure can be used for high resolution remote sensing image feature extraction.As to the complex characteristics of remote sensing images,with analysis of feature extraction,a model called joint-layer deep convolutional neural networks is proposed for object detection in high resolution satellite remote sensing images.The main difficulties of object detection in remote sensing images are complex and changeable background,relatively small object and unobvious feature.The combined features can detect local features of remote sensing images easier.Joint-layer features are applied to train support vector machine classifier for vehicle detection at last.Experimental results show that joint-layer deep convolutional neural networks improve the precision rate by 16% and the recall rate by 6% compared with traditional convolutional neural network in vehicle detection,let alone outperform traditional artificial feature methods.It demonstrates that the proposed model is effective for the feature learning of small objects in high resolution satellite remote sensing images.Finally,we combine convolutional neural network and recurrent neural network to train a multi-modal language model,which is used to solve the semantic problem of high resolution satellite remote sensing images.Natural datasets are added to remote sensing datasets to enhance the generalization performance.In the training stage,convolutional neural network is used as an encoder to extract image features.The image features and word vectors of the corresponding sentence are mapped to feature space in the same dimension.Gradient descent algorithm is applied to train the generative recurrent neural network model.Sentences predicted by the model are evaluated through BLEU automated metric;in the testing stage,images are put into the generative recurrent neural network model to produce image caption.The experimental results show that the multi-modal recurrent neural network model achieves certain results in remote sensing image caption.However,limited by the size of datasets,the accuracy of semantic descriptions can be improved by increasing remote sensing data.