On The Remote Sensing Object Detection Algorithm Based On Rotationally Anchor-free Bounding Box

Remote sensing object detection algorithm is a new research direction in the field of object detection in recent years,and has received more and more attention.Compared with ordinary object detection scene,remote sensing scene has the characteristics of large scale change of objects,many small objects,disordered arrangement of objects and complex background environment,which make it difficult to detect accurate objects.This paper combs the development of object detection algorithm based on neural network,analyzes the advantages and disadvantages of existing algorithms,makes a detailed study on various algorithms of remote sensing object detection.And finally selects the object detection anchor-free algorithm and rotating box as the design direction,and designs a new remote sensing object detection algorithm.First of all,I studied the main existing anchor-free algorithms and analyzed their advantages and disadvantages.Combined with the characteristics of remote sensing environment,I selected CenterNet as the benchmark network to design.In this paper,the network structure and loss function of CenterNet are deeply studied.And the excellent design is retained,the improvement is added in the follow-up.Then the common oriented box is studied.The characteristics of the basic five parameter oriented box and eight parameter oriented box are analyzed.Then the paper studies the evaluation index map(average precision)used in object detection,and analyzes its rationality as the evaluation standard.Secondly,on the basis of the previous part,I made in-depth research on the existing design of various deformation of the oriented box.Based on the improvement of the eight parameter oriented box,I designed the multi vector regression box,and designed the supporting loss function.Combined with the use,I can better regression the rotating object in the remote sensing field.The geometric relationship between the regression box and the prediction vector can be used to constrain the prediction results,and the relationship between the vertices of the regression box can be used to avoid the problem that the distance loss function can only be optimized separately for a certain coordinate,so as to make better use of the global information.Then,experiments are carried out on the commonly used remote sensing dataset UCAS-AOD,and the results are compared and analyzed to prove the effectiveness of this part of the design,and ablation learning is carried out to prove the necessity of collinear loss function,vertical loss function and overall width height loss function.Thirdly,the common spatial attention and channel attention blocks are studied,and the network structure is designed.The attention mechanism is combined in the specific location of Res Net network,so that the network can better focus on the possible location of the object and reduce the interference of background information.Then combined with the previous multi vector regression box and loss function design,complete the overall design of the algorithm.A large number of experiments are carried out on three commonly used remote sensing object detection datasets.The comparison results prove the universality of the algorithm I designed.And compared with other excellent algorithms,the effect is better than the best anchor-free algorithm in the field of remote sensing object detection.Finally,the shortcomings of the network are analyzed,and a variety of well-known algorithms using multi-scale information are studied to improve the network structure.The backbone network is replaced by resnext network which uses a variety of convolution kernals in parallel.The deformable convolution is introduced into the network structure to change the sampling points of convolution operation.And the sampling points are distributed to the regions with richer features,.So that the network can obtain information of different receptive fields and learn multi-scale features.Experiments on UCAS-AOD dataset show that the optimization of network is effective.