Research On Anchor-Free Object Detection Technology In Remote Sensing Images Based On Deep Learning

In recent years,with the continuous improvement of computing ability of computers,convolutional neural networks and deep learning technology have been widely used in different scenes.On the one hand,in the field of computer vision represented by object detection,the use of deep neural network can greatly improve the detection performance;on the other hand,with the development of drone and remote sensing technology,the importance of remote sensing image analysis is constantly reflected,and remote sensing image analysis has become a hot spot in both military and civil fields.As a basic task in the field of computer vision,object detection can accurately capture the target and provide rich and accurate information for subsequent tasks.Object detection has been widely and deeply studied in natural images,and the detection performance is constantly improved in the works using deep learning technology.However,remote sensing images have special problems compared with natural images.Firstly,as remote sensing images are usually taken from the perspective of bird’s-eye view,the target is always arbitrary-oriented.Secondly,there are various categories of objects in remote sensing images,and the size of different objects varies greatly due to the shooting height.Thirdly,the target in remote sensing images is located in variable environments,and the complex backgrounds bring difficulties for detection.Finally,in the object detection systems using convolutional neural networks,the general convolution operation is not aligned with the feature extraction of oriented objects.The performance is usually far from satisfied if simply apply the general object detectors to the detection of remote sensing images.Therefore,it is necessary to design the detector according to the special problems in remote sensing images.This paper focuses on the above three points,and carried out from four aspects.The main works are included as follows:(1)We design a baseline model considering the problem of arbitrary-orientation of objects.To eliminate ambiguity,an indirect oriented bounding box representation is designed to obtain the angle information,which decomposes the target bounding box regression into two sub-tasks:the horizontal object bounding box regression and the coordinate offsets regression.At the same time,we build an"anchor-free" baseline model which avoid complex design of anchors.(2)For the problem of scale variation,this paper design a finer adaptive feature fusion network based on feature pyramid networks to obtain more abundant and accurate feature information,which reduces the effects of course assignment in feature pyramid networks and improves the overall performance of the detection model.(3)For the problem of background noise in remote sensing images,we first combining the supervised pixel attention network and channel attention network as foreground enhancement network to make the feature maps learned by neural network focus more on object instances.Then,a new sampling strategy is designed,which can obtain high-quality samples and distinguish the foreground from background effectively.(4)For the problem of feature misalignment resulting from arbitrary orientation of objects in general convolution feature extraction,an angle adaptation module is designed based on deformable convolution.Multiple branches with different angle offsets are constructed and weighted by pixel attention network.The proposed angle adaptation module can enhance the robustness of the detector.