Research On Object Detection Method In Remote Sensing Image

Remote sensing images cover a wide range of targets,including man-made objects(such as vehicles,ships,buildings)with sharp boundaries and independent of the background environment,and landscape objects(such as mountains,hills)with blurred boundaries that are part of the background environment.In addition,it has broad application prospects in the fields of precise attack of military targets,national defense security,land resources investigation,ecological environment monitoring,natural disaster monitoring,and geological disaster detection.Object detection in remote sensing image is the core technology of remote sensing image processing and analysis.It aims to detect whether a given remote sensing image contains one or more interested objects and determine the location of each object in the image.However,due to such factors as line of sight angle,target occlusion,complex background,illumination intensity,and shadow graphics,the visual appearance of the object will change greatly,which will lead to the difficulty of detection.Therefore,how to quickly and accurately detect and extract the typical objects in various complex environments in remote sensing images,and improve the accuracy and efficiency of automatic detection of remote sensing image data has become an urgent technical problem,and has important academic value and engineering significance.With the breakthrough of deep neural network technology,the method of feature extraction and representation based on convolution neural network has achieved great success in the field of natural image object detection.However,due to the difference between remote sensing image and natural image,this kind of method cannot be directly applied to remote sensing image object detection task.Therefore,in view of the above challenges,this paper takes the method based on deep learning and convolution neural network as the breakthrough point,and combines the characteristics and difficulties of remote sensing image to carry out in-depth research.First of all,to solve the the problem of poor detection performance in remote sensing images,a new multi-scale fusion region proposal detection framework based is proposed.In the region proposal generation stage,a multi-scale feature pyramid structure is constructed through multi-model fusion,which can effectively generate the region proposal blocks with a wider scale range.In the region proposal classification stage,an improved image block-voting algorithm is proposed.By adding dynamic weights to the generated region proposal,the background block interference is suppressed,and the object image block is enhanced and highlighted,thus effectively improving the classification accuracy of the model.In the post-processing stage,an effective overlapping de-duplication algorithm is proposed,which further improves the detection performance of the model.In addition,due to the lack of common data sets of building objects,a new data set containing different types of buildings is constructed,which aims at training models and verifying the performance of models in building detection tasks.Finally,the effectiveness and robustness of the proposed multi-scale detection model are verified by experiments.Secondly,to solve the problems of confusion in multi-class object detection and poor performance in small target detection in remote sensing images,a detection framework based on scene-contextual feature pyramid network is proposed.In the proposed framework,considering the characteristics of the object generally appear in the specific scene in remote sensing images,it constructs a novel network structure that integrates the scene context features with the object features.Based on the feature pyramid framework,which has good detection performance for multi-scale targets,this structure effectively improves the classification accuracy of multi-class objects by establishing and enhancing the relationship between scene and object,thus improving the detection accuracy.On this basis,the structure of block of the backbone network is further improved by combining residual block with dilated convolution,and forming a novel block structure.Compared with the ordinary residual block structure,the proposed block structure has stronger feature extraction ability and is more sensitive to small object detection,which is helpful to improve the detection accuracy of the model.In addition,considering the large size of remote sensing images,using batch normalization will make the training model consumes too many resources.Therefore,we introduce the group normalization for the first time,which divides the channels into groups and computes within each group the mean and variance for normalization,for solving the limitation of the batch normalization.Finally,to solve the problem that the existing detection methods are not effective in dense and small objects detection in remote sensing images,a novel non-maximal suppression(NMS)method with mask-intersection-of-union(mask-Io U)is proposed.In the proposed method,a dynamic-threshold NMS method,which based on the NMS method,for detection boxes de-duplication by calculating the Io U of mask area of the objects,and achieves excellent performance on the dense object detection in remote sensing images.In addition,due to the irregularity of the area of mask,it is difficult to obtain the mask-Io U by conventional methods.Therefore,in the processing of calculating the mask-Io U,an approximate calculating rule is proposed,which can effectively obtain the mask-Io U by approximate regularization of the mask.