Research On Deep Learning Object Detection Methods Based On High Resolution Remote Sensing Images

With the development of high-resolution imaging technology in China,the difficulty of obtaining high-resolution Remote Sensing Images(RSIs)is greatly reduced,and the spatial resolution is greatly improved.These changings bring a large number of high-resolution RSIs with complicated contents.It is difficult to deal with the complicated images quickly,automatically and effectively based on the traditional manual features.In recent years,the deep learning-based algorithms get developed and can adaptively learn and deal with RSI complex features and achieve excellent performance.Therefore,deep learning is widely used in various image processing fields,such as scene classification,object detection,and semantic segmentation.Among these fields,the object detection task aims to determine the position and category of the object from the input image.The RSI object detection is widely used in many military,civil,economic and other fields,thus shows great importance.However,due to several reasons,it is hard to implement deep learning algorithm into RSI object detection.1)the object spatial deformation,such as rotation and scaling in RSIs;2)RSIs contain a large number of small objects which is difficult to detect;3)the RSI content is complex and confusing;4)the region proposal is rough.Therefore,according to these characteristics and problems of high resolution RSIs,how to build a deep learning object detection system for high resolution RSIs is the key to the research of high resolution remote sensing object detection.In this paper,the research is based on the deep learning object detection system,aiming at the characteristics of high resolution RSIs(such as object spatial deformation,the large number of small objects,confusing objects,and rough regional proposal),different problems in the system are studied.1)feature extraction,2)object localization,3)object recognition and 4)localization-recognition coupling.Therefore,the deep learning object detection system is expanded and updated in the field of high resolution RSI analysis.1)For the feature extraction of the detection system,in view of the spatial deformation characteristics of the RSI objects,such as rotation and scaling in high resolution RSIs,this paper proposes a Hierarchical Robust Convolutional Neural Network(HRCNN).This framework combines the advantages of fully connected features for spatial shape change which is conducive to object recognition,and the hierarchical spatial semantic information of convolution features which is conductive to object localization.The network can effectively improve the expression ability of features for the image content.In addition,a large,13-class and different categories object instances balanced high-resolution remote sensing object detection data set is established to train the deep learning object detection model.2)For the problem of object location in the detection system,aiming at the characteristics of containing a large number of small objects that are difficult to detect in high resolution RSIs,a Gated and Axis-Concentrated Localization Network(GACL Net)is built.This framework proposes to map the convolution feature map to the horizontal and the vertical axes,and correspondingly predicts the horizontal and vertical coordinates of the object bounding box separately,so as to avoid mutual interference between the horizontal and the vertical coordinates prediction.The fully connected feature is used to guide the convolutional feature with global semantic information through a channel attention mechanism,in order to make up for the lack of semantic information in convolutional features.According to the organic combination of the above two modules,the proposed model can effectively improve the performance of the detection system in object positioning.3)For the problem of object recognition in the detection system,aiming at the characteristics of a large number of confusing visual fine-grained categories in high-resolution RSIs,an Attribute Cooperative Convolution Neural Network(ACCNN)is built.This framework proposes to use collaborative learning between the attribute task and the classification task,to provide more abundant information for the classification(recognition)task,and increase the ability to distinguish the hard and visually fine-grained categories.Moreover,a relationship learning tasks is established between the classification task and the attribute task.It can further enhance the information sharing between the classification and the attribute tasks,and enhance the ability of classification branches.The proposed framework can effectively reduce the error rate of recognition.4)For the coupling problem of location and recognition in detection system,aiming at the coupling error transmission problem caused by rough region proposals of high resolution RSIs,a Consistent Multi-Stage Detection(CMD)network framework is built.This framework proposes a new network head with a novel coupling method between classification and recognition,which is suitable for rough input proposals;while the normal network head with traditional coupling mode can obtain more efficient performance in the case of fine input proposals.Therefore,in this paper,through the study of the two types of coupling network heads,to be consistent with the changing trend of the proposals from coarse to fine,this paper puts forward the proposal-consistent network heads cooperation mechanism between the two network heads.The proposed cooperation mechanism gives full play to the advantages of the two network heads with different coupling mode,so that their advantages can complement each other.