基于多分支特征对齐的遥感图像旋转目标检测网络

A Multi-branch Feature Alignment Network for Arbitrary-oriented Object Detection in Remote-sensing Images

具有高分辨率、多尺度、多方向等特点的遥感图像目标检测是一项重要且极具挑战性的工作.近年来,基于深度学习的研究方法取得了良好的进展,同时针对遥感旋转目标检测的工作也有了显著的发展.在自然图像方面,水平目标检测方法的性能已经比较成熟,然而使用该方法检测遥感旋转目标难以获得理想性能.基于水平向模型设计了旋转目标检测方法,并提出了一种多分支特征对齐网络(MBFA_YOLO)来提升检测性能.具体来说,保留基线模型的结构,在骨干网络上使用并行分支得到不同尺度的浅层特征,利用不同膨胀系数的复合膨胀残差模块(D_Resnet)对其进行学习,然后针对分类任务和回归任务设计不同的注意力机制.考虑到现有的单级检测器中特征错位的不足,设计了一个特征对齐模块来获得更准确的特征.仿真实验结果验证了该方法的有效性.

Object detection in remote-sensing image with high resolution,multi-scale and multidirection characteristics is an important and challenging task.In recent years,the methods based on deep learning have made good progress,and the detection of remote-sensing oriented objects has also made significant development.In terms of natural images,the performance of the horizontal object detection methods is relatively mature,but it is difficult to obtain the ideal performance by using this methods to detect remote-sensing oriented objects.In this article,a oriented object detection method is designed based on the horizontal baseline model,and a multi-branch feature alignment network(MBFA_YOLO)is proposed to improve the detection performance.Specifically,this paper retains the structure of the baseline model.This paper proposes a method to obtain shallow features with different scales on the backbone network by using parallel branches,and uses the composite dilated residual module(D_Resnet)with different dilated coefficients to learn them.Then,different attention mechanisms are designed for classification and regression tasks.Considering the shortcomings of feature misalignment in existing single-stage detectors,this paper designs a feature alignment module to obtain more accurate features.Experimental results validate the effectiveness of this method.