关键词: 计算机应用, 遥感目标检测, Meta-ACON激活函数, 加强连接特征金字塔网络, 角度预测, 旋转检测框

Abstract:

Due to the dense arrangement and different directions of objects in remote sensing images， the existing detection algorithms are difficult to locate the instance objects accurately. Therefore， this paper proposed an omnidirectional accurate detection algorithm for dense small objects in remote sensing images. In order to improve the ability of feature extraction， the meta-ACON activation function was introduced into the residual structure of the backbone network to adaptively learn the importance of channel features. A new strengthen tconnection feature pyramid network was proposed. The lateral connection part for deep and shallow feature fusion was redesigned， and a jump connection was added between the input and output of the same level feature map to enrich the feature semantic information. The angle prediction branch was introduced， and the circular smooth label method was used to transform the angle regression problem into a classification problem. While realizing the rotation of the object frame， the problem of sudden change of the boundary of the rotation frame was solved. A post-processing method （Rotate-Soft-NMS） for rotation detection frame was designed to remove adjacent repeated rotation detection frames by suppressing the confidence of the detection frame. The experimental results on DOTA dataset show that the mAP of the proposed algorithm is 76.15%， which is 5.22% higher than the benchmark model YOLOv5m， and has achieved the best detection results compared with other advanced algorithms. The algorithm in this paper has achieved better detection results for complex remote sensing images.

Key words: computer application, remote sensing object detection, Meta-ACON activation function, strengthen connection feature pyramid network, angle prediction, rotation detection box