Airplane Target Recognition In Remote Sensing Images Based On Convolutional Neural Networks

With the rapid development of modern remote sensing technology,remote sensing images(RSIs)have become an indispensable data resource and are widely employed in the civilian and military field.Airplane target recognition in RSIs is one of the important research directions of RSI interpretation and analysis,which has received a great deal of attention in academic and industrial circles.The higher the spatial resolution of RSIs is,the richer information the image includes.Utilizing traditional manual interpretation to recognize airplane target in RSIs has been unable to meet the requirements of social development.Therefore.how to extract the information about the locations of the airplane and even the types of the airplane accurately and efficiently from RSIs is the emphasis and difficulty of RSI interpretation.Deep learning is becoming an emerging technique with the advent of ear of big data in recent years.The convolutional neural network(CNN),as one of the most representative deep learning models,have made revolutionary achievements in a wide range of fields,compared with traditional methods.Based on the achievements of CNNs in the computer vision field,in this paper,we center on the application of CNNs in airplane target recognition in RSIs,and focus on two aspects:airplane detection and airplane type recognition.The main work and contribution of this paper are summarized as follow:(1)Nowadays,the existing object detection methods based on CNNs require a large amount of training images with manual annotation.such as labeling the locations of each object.However,in RSIs,manual annotation of airplane targets is generally expensive and sometimes unreliable.To tackle this problem,we propose a novel airplane detection method in remotes sensing images based on weakly supervised learning.The proposed method makes use of the CNN to detect airplane targets in RSIs only requiring the training images is annotated with the labels that indicate whether the image contains the airplane targets or not whereas locations are not provided.The experimental results show that our proposed method can perform comparably to the method using airplane location annotations for training and achieve a low false alarm rate.(2)Aiming at improving the efficiency of the traditional image feature representation and recognition model of airplane target,in this paper,we employ the CNN to recognize the types of airplane in RSIs.However,there are no open source and mature datasets concerned with airplane type recognition in RSIs,and the training images of the specific types of airplane targets are not enough.Therefore,in this paper,we establish a new dataset.11 Types of Airplane in Remote Sensing Images(ARSI-11).Based on this dataset and the idea of transfer learning,an airplane type recognition method in RSIs is proposed.The proposed method effectively realizes the airplane type recognition in RSIs by utilizing the powerful capabilities of learning and feature representing provided by CNNs.We present a detailed experimental evaluation on the ARSI-11 dataset.which verifies the effectiveness of our methods.