Airport Detection In Optical Remote Sensing Images With Convolution Neural Network

The existing target recognition methods in remote sensing images cannot identify the key target fast and accurately.However,in the field of natural images,the target recognition methods based on convolutional neural network has achieved great success in recognition precision and speed.In this paper,we make researches on improving the accuracy and speed of airport recognition in optical remote sensing images,learning from the object recognition methods in natural images.This paper analyzes the shallow visual features of the airport,including intuitive features and handcrafted features.Then an airport detection method based on shallow visual features is proposed,which uses information of line segments to generate candidate windows and extracts the texture and SIFT features into SVM classifier for image classfication.This method improves the recognition accuracy by several percentages.In view of the improvement in speed of target recognition in natural images made by Faster RCNN,this paper applys this method to detect airports in optical remote sensing images with some modifications.On account of the commonness between image features of natural images and remote sensing images,we transfer the feature representation from a pre-trained network model for airport recognition.Experimental results show the effectiveness of network transferring and the airport identification is accelerated.To improve the region proposal method,we propose another airport recognition algorithm based on prior information and convolution neural network.The candidate windows are generated from line segment sets and filtered by the prior information,which includes the length and width of the runways and the distribution of line segments.By this method,the number of candidate windows is greatly reduced and then Alexnet is used for airport recognition.Experimental results show that the method make progress in recognition accuracy and speed.