| Remote sensing object detection is the focus of remote sensing image processing and plays an important role in aerospace,military and civil fields.However,it is still a great challenge to realize the effective detection of multi-scale remote sensing objects densely arranged in a complex background for the real-time extraction of object information in orbit from highresolution remote sensing images.In this paper,through the analysis of the characteristics of high-resolution remote sensing object,considering precision and speed,the existing object detection algorithm is improved from the aspects of the multi-scale feature fusion,feature enhancement,context information supervision and model compression,a new detection network EFEC-MNet is proposed.While ensuring the detection speed,the detection accuracy is improved,which provides more ideas for the future research in the field of remote sensing image object detection.The research contents and main contributions of this paper are as follows:(1)Considering the long shooting distance of high-resolution remote sensing image,wide field coverage and large image size,it will consume a lot of computing resources to input such image data directly into the network for training.If the image is scaled to the size suitable for network input,the information of small object will be lost in the convolution process.In this paper,the original image is splitted into 640*640 by using the image split strategy with overlapping.The experiment shows that the detection precision of the splitted mAP is increased by 22.22% compared with that without image split.(2)For the data characteristics of high-resolution remote sensing objects diversity of scale and more small objects,this paper proposes a MF-Net based on multi-scale feature fusion to improve the detection accuracy based on the optimization of the existing detection network.Taking the hourglass network as the backbone of detection network,the feature fusion of the low and high levels of these scales is carried out in the network,and the corresponding detector is input to realize the effective detection of the four scale objects.Compared to the YOLOv3 network,the mAP of the MF-Net increased by 2.62% on the RDSD dataset,and the detection speed increased by 1 FPS.(3)In order to further improve the detection accuracy of tiny objects,this paper introduces the feature enhancement residual structure(FERM)into MF-Net,the MFE-Net based on feature fusion and feature enhancement is proposed.By increasing the receptive field of the feature map,the detail features of the tiny object are enhanced,so as to improve the detection accuracy.Soft-NMS is used for the post-processing of the boundary box,which improves the detection accuracy of objects densely arranged to some extent.MFE-Net has the same detection speed as YOLOv3 network,and mAP improves by 4.15%.(4)In order to solve the problem of mis-detection and omission of objects under complex background,this paper introduces the stacked hourglass network(SHN).By combining the context information of the feature map,using an intermediate supervision strategy,and involves the output features of two levels into the loss calculation at the same time,and FECMNet based on context information supervision is proposed.Compared with MFE-Net,the mAP of FEC-MNet improved by 2.92%,but the detection speed decreased.In order to improve the detection speed of FEC-MNet,this paper proposes a compression model EFECMNet based on asymmetric convolution.Compared with FEC-MNet,the detection speed of EFEC-MNet increased by 36.11%,and 72.67% mAP was obtained on the DOTA dataset. |
