Multi-class Object Detection Technique For Complex Large-field Optical Satellite Remote Sensing Images

With the rapid development of remote sensing technology and the improvement of sensor imaging resolution,object detection based on remote sensing images has become one of the important research directions in computer vision(CV).In recent years,due to the breakthrough of hardware computing power,the target detection method based on deep learning has ushered in a blowout growth,and this technology has also been brought to the field of remote sensing by researchers.However,due to the large field of view remote sensing images often have complex backgrounds,large differences in target scales,dense arrangement of small targets,and arbitrary target orientations,the currently commonly used general-purpose detectors based on deep learning perform poorly on remote sensing images.Some targeted improvement networks have improved their performance on remote sensing images to a certain extent,but the above problems still exist.Therefore,this research constructs a detection technology for multi-type targets in complex and large-field optical satellite remote sensing images.The main innovations include a feature optimization network based on3 D attention and recursive multi-scale,and a rotating target detection method based on dynamic anchor matching.two parts.The key contents are as follows:(1)In view of the complex background of remote sensing images,the large difference in scales of various targets,and the dense arrangement of a large number of small targets,the existing methods do not fully mine feature information in the feature extraction stage.A target feature optimization network for multi-scale feature fusion.Among them,the three-dimensional attention module mines the importance of each neuron by optimizing the defined energy function,and derives the three-dimensional attention weights for the feature map without any parameters,which enhances the target feature representation in complex backgrounds,especially The characteristics of small targets can suppress the noise information brought by the background and nearby targets.The recursive multi-scale feature fusion network feeds the output of the FPN into the bottom-up Res Net backbone network,and adopts ASPP at the interface to improve the receptive field and multi-scale performance of the feature map,and then fuses two recursion through the fusion module.The output of the operation promotes the recursive and deep fusion of low-level texture information and high-level semantic information,achieves more robust feature extraction,and lays a feature foundation for the subsequent high-precision detection of remote sensing targets.(2)In view of the particularity of the bird's-eye view angle of optical remote sensing images,the remote sensing targets are densely arranged and have arbitrary directions,and the currently commonly used detection methods based on horizontal boxes cannot be adapted.In this paper,a rotating target detection method based on dynamic anchor matching is constructed..Specifically,it includes the design and allocation strategy of Anchor,the solution to the boundary problem in the rotation box representation,the design of the loss function and the improvement of NMS.First,a better anchor assignment is obtained by dynamic anchor selection based on matching degree(MD).Secondly,the RDIOU-Smooth-L1 rotation box regression loss function is proposed,which solves the boundary problem in the rotation box representation,and the phenomenon of misalignment between the smooth L1 loss in the regression and the IOU-based evaluation index.In addition,the design of the classification loss combines the Focal-loss and the matching compensation factor,which balances the positive and negative,difficult and easy samples and distinguishes the positive samples with different positioning potentials.Finally,by introducing the center point distance and direction information of the detection frame in the NMS stage,the RDIOU-NMS is constructed,which effectively retains the adjacent target frames and eliminates duplicate detection frames for targets in any direction and densely arranged.In this paper,ablation experiments and comparative experiments are carried out on the large-scale public remote sensing image datasets DODA-v1.0 and HRSC2016,respectively,to verify the above modules and the overall method,which proves the effectiveness of the above modules and the overall method.