Fast Detection Of Aircraft Target In Large-Scale Optical Remote Sensing Image

With the rapid development of high-resolution optical satellite technology,remote sensing satellite imaging resolution has been greatly improved.The number of optical satellites has rapidly increased,and revisit cycle has been greatly shortened,resulting in a sharp increase of high-resolution image data.Massive high-resolution optical remote sensing image data based on Earth observation satellites has been applied in different research fields for the purpose of military safety and civil service.The task of automatic detection and recognition of aircraft has always been a hot research topic and an important part in aerospace fields,which plays an essential role and have strategic needs in reconnaissance,early warning and other areas of military defense.The timeliness of traditional target detection technology is difficult to adapt the numbers of high-resolution remote sensing image data,the features which rely on were artificially designed,time consuming and strongly dependent on the expertise and characteristics of the data itself.In addition,targets are far from the field of view and in proportion are limited in remote sensing images,such methods for the detection of small targets is still weak.Moreover,due to the multi-temporal nature of illumination,imaging angle and weather,a small amount of even missing texture and edge information will bring great difficulties and challenges to the detection and recognition of the target.The convolutional neural network technology has inherent advantages incomparable to the traditional methods in the deep feature extraction of the image.At the same time,the theory and technology in Deep-Learning made great progress,the related algorithms of target detection and recognition have been rapidly developed,promoting the application performance in fields of computer vision,providing new ideas for the interpretation of remote sensing data.Thus,we focus on the research of high accuracy detection and recognition of aircraft targets in large-scale high-resolution optical remote sensing data under the premise of ensuring the detection speed.The main body of our research can be summarized into the following three pieces of work:1.Remote sensing image enhancement and hazing removeIn order to reduce the impact of the camera's physical environment and illumination on image quality during imaging,the image enhancement algorithms such as remote sensing image noise reduction and illumination homogenization are sequentially studied.In addition,the missing details of the target which to be detected would be caused by the smoke or fog that brought by the multi-temporal nature of the earth climate and weather during imaging of Earth Observation Satellites,the remote sensing images are enhanced based on the algorithm of image haze removal by Dark Channel Prior.2.Research on remote sensing image airport detection algorithmIn order to reduce the detection field,lock the detection range and reduce the computational complexity of target search,it is necessary to identify and calibrate the airport area in the image.The Hof line detection method is adopted and the convolution classification network is used to detect the airport runway to locate the airport area,so as to lay the foundation for subsequent aircraft recognition.3.Research on the algorithm of aircraft target detection in remote sensing videoIn order to carry out fast target detection of large-scale remote sensing video images,and considering both the detection speed and the precision,this paper proposes a fast target detection algorithm based on You Only Live Once(YOLO)for large-scale remote sensing images.Compared with the original version of YOLO and YOLOv2,the improved algorithm has a larger detection field of view and detection precision under the condition of guaranteed detection speed.