Object Detection In High Resolution Remote Sensing Imagery Based On Convolutional Neural Network

With the development of earth observation satellite technology,high resolution remote sensing satellite has become an important part of high resolution earth observation system.Object detection using high spatial resolution remote sensing images(HSRIs)is a fundamental method for image information automatic extraction,analysis and understanding,and plays an important role in both civilian and military applications.Object detection in HSRIs is to determine whether a HSRI contains one or more objects belonging to the class of interest and locate the position of each predicted object in the image.The term,‘object’,used in this paper mainly refers to man-made objects(e.g.,airplanes,storage-tanks,or ships)that have clear boundaries and are independent of the image background.The majority of the object detection methods typically adopt a three-stage mode of(1)extracting object candidate regions,(2)obtaining object candidate region features,and(3)feature classification of the object candidate regions to realize image object detection.There are complex scenes in HSRIs.The three-stage mode is difficult to adapt to the object detection of a large number of HSRIs under complex conditions.The robustness and universality of the three-stage mode are poor.In recent years,deep learning has attracted the attention of different scholars.Convolutional neural network(CNN)models are the most popular deep learning models.CNN does not require the use of handcrafted features,and can extract and learn effective features itself according to massive data and annotations.Moreover,in the case of sufficient training data,CNN has acceptable generalization ability and can maintain reasonable robustness under complicated and variable conditions.Therefore,object detection framework based on CNN has been widely applied to object detection in HSRIs,such as Faster-RCNN,single shot multibox detector(SSD)and you only look once(YOLO).However,these object detection frameworks based on CNN are designed for natural images.Compared with natural images,HSRIs have the characteristics of more complex background,smaller object scale,larger scale change of similar objects and larger image size.When them are directly used to detect objects in HSRIs,there are some problems such as the mismatch between the regional proposed anchor scales and the HSRIs object scales,the size of HSRIs is too large relative to the networks input,and lacking training data.With respect to above problems,a large-scale HSRIs object detection data set is established.The suitable regional proposed anchor scales of object detection framework based on CNN are obtained by compiling statistics for the scale range of objects in HSRIs object detection data set.The optimal block method for object detection in largescale HSRIs is obtained by studying to use different block to detect objects in largescale HSRIs.Through the research in the paper,it provides theoretical and technical support for object detection in large-scale HSRIs.The research content and innovation of this paper is summarized as follows:(1)Object detection in high resolution remote sensing imagery based on convolutional neural networks with suitable object scale featuresHSRIs are obtained by satellite or space sensors in near-earth orbit using a topdown view.The objects are relatively small in large-scale HSRIs and appear in densely distributed groups.Therefore,how to design a object detection framework based on CNN suitable for small-scale object detection is crucial for obtaining high-precision HSRIs object detection results.A large-scale HSRIs object detection data set is established.The suitable regional proposed anchor scales of object detection is obtained by compiling statistics for the scale range of objects in dataset.Then,a CNN framework for object detection in HSRIs is designed using suitable regional proposed anchor scales of object detection.The research content can well solve the problem of small object scale and large change of similar object scale in HSRIs.(2)Multiscale block fusion object detection method for large-scale highresolution remote sensing imageryHSRI products are typically large-scale,such as the size of Gao Fen-2(GF-2)is29,200 × 27,620 pixels.When an object detection framework based on CNN is applied to large-scale HSRIs,it is necessary to divide the HSRIs into blocks.In this paper.Frist,objects in large-scale HSRIs are detected using different block scales and the average precision(AP)of the different object detection results are counted at different block scales.Then,according to the statistical information,the image block scales corresponding to the optimal AP value of the different objects are obtained.Finally,the image block scale object detection results corresponding to the optimal AP values of the different objects are fused using a soft NMS(Soft-NMS)algorithm on the object detection results of the large-scale HSRI.The research content provides an optimal block strategy for object detection in large-scale HSRIs.(3)Object detection method for high resolution remote sensing imagery using convolutional neural networks based on embedded GPUWith the development of earth observation technology,the requirement of on-orbit image object detection becomes more urgent.Embedded GPU has the characteristics of small size,low power consumption and fast operation.It can be easily mounted on various on-orbit processing platforms.In this paper,a convolution neural network based on embedded GPU for object detection in HSRIs is proposed to realize on-orbit HSRIs object detection.Through the way of satellite-ground cooperative processing,the parameter model of object detection framework based on CNN is trained in ground server,and the object detection framework based on CNN and parameter model are transplanted to embedded GPU to realize HSRIs on-orbit object detection.The research content provides theoretical and technical support for on-orbit object detection in HSRIs using convolutional neural networks based on embedded GPU.In view of the above research content,two large-scale target detection data sets WHU-RSONE and WHU-RSONE2.0 are used to qualitatively and quantitatively compare and evaluate the of this research content.The experimental results demonstrate that the proposed method can effectively couple the scale range of various objects in HSRIs,and fully use the advantages of each single-scale block object detection,and compensate for the disadvantage of the single-scale block object detection,where it can difficult to detect certain objects in an image.The object detection results of large-scale HSRIs can be accurately obtained using proposed method.The proposed method is transplanted into embedded GPU to realize on-orbit image object detection using object detection framework based on CNN.