Airport And Airplane Detection In High-Resolution Remote Sensing Imagery Based On Convolutional Neural Network

Typical object detection in high-resolution remote sensing imagery(RSI)has always been an important part of remote sensing image interpretation.With the rapid development of satellite remote sensing imaging techniques,how to extract the required information from RSI and detect the specific target quickly and accurately is the hot spot of current research.Traditional methods of object detection generally take the idea of extracting the hand-crafted features to train the classifiers.Therefore,how to choose the representative and discriminative features of objects has become a key factor to improve the detection accuracy.Convolutional Neural Network(CNN)is a kind of deep learning model,it can learn the features from images automatically by establishing a hierarchical structure which is similar to human brain.At the same time,it also has the ability to deal with big data.In recent years,CNN has been applied to face detection,speech recognition and image classification successfully,but it is seldom applied to remote sensing fields.Hence,exploring to apply CNN to remote sensing field has important theoretical significance and broad application prospects.This paper focuses on the application of CNN in object detection in remote sensing imagery.Considering of the characteristics of the airports and airplanes,we propose different algorithms based on CNN to detect the airports and airplanes in RSI.The main research contents and innovations in this paper can be described as follows:(1)An airport detection algorithm based on visual attention computation model and CNN-BoW model is proposed.Firstly,GBVS model is used to detect the salient regions in RSI,and the regions of interest of airport(ROI)are extracted according to the saliency map.Secondly,the Hough transform is used to detect the lines in the ROIs for preliminarily screening and optimizing the ROIs,which can reduce the computation cost and avoid the interference caused by the large area of background.Finally,the features of ROIs are obtained by CNN-BoW model,and the ROIs are classified by SVM classifier.The classification results of the airport areas are labeled in the original remote sensing image as the final detection results.In CNN-BoW model,the features extracted by CNN are taken as words to construct visual dictionary for training the BoW model,in this way,the images can be described by features extracted by CNN-BoW model.A large RSI airports dataset is set up for airport detection.The experimental results show that the words from CNN are more discriminative,which improves the detection accuracy of the proposed airport detection method.Compare to the SIFT features,the proposed airport detection method can not only improves the recall rate about 10%,but also reduces the false alarm rate about 8%.(2)A novel CNN framework called Multi-Struct Convolutional Neural Networks(MSCNNs)is proposed to detect the airplanes in RSI.Four CNNs with different network structures are designed by considering the size and the number of convolution filters,and the number of layers in CNN.In order to fully utilize different features learned by these four CNNs with different structures,the outputs of them are concatenated as a kind of new features which are more comprehensive.Then these new features are put into a SVM classifier to fulfill the following classification task.The object detection experimental results on a RSI airplane dataset show that MSCNNs has obvious advantages than any single-structure CNN.The average recall rate is improved about 5%,while the average false alarm rate is reduced about 8%.