基于CAFPN和细化双头解耦的遥感图像目标检测

Remote Sensing Image Object Detection Based on CAFPN and Refinement Double-Head Decoupling

针对遥感图像背景的复杂性和图中目标尺寸小、方向任意性导致漏检或错检的问题,提出了一种新颖的目标检测算法。首先,提出一种基于上下文信息增强的特征金字塔网络。在特征提取阶段,自适应融合不同感受野,获得具有丰富语义信息的特征,减少小目标的信息流失。然后,在回归网络中,使用中心点偏移回归机制实现旋转框的检测,降低冗余锚框带来的计算复杂度。最后,结合双头网络将分类和回归特征解耦,通过注意力机制和极化函数引导的特征细化模块构建适应各自任务的重要特征,使网络能准确地检测目标。在遥感数据集DOTA、HRSC2016和UCAS_AOD上验证网络的有效性,对比于Faster RCNN算法,该算法在3个数据集上获得了8.48%、7.60%和3.10%的精度提升,实现了高性能的遥感图像目标检测。

In order to address the challenges posed by complex backgrounds,small object sizes,and arbitrary directions in remote sensing images,this paper presents a novel object detection algorithm.The proposed algorithm consists of several key components.Firstly,a context augmentation feature pyramid network(CAFPN)is introduced.In the feature extraction stage,it integrates adaptively with the dilated convolution to obtain features with rich semantic information and reduce information loss of small objects.Then,midpoint-offset regression(MOR)is employed to detect oriented box in the regression network to reduce the computational complexity caused by redundant anchors.Finally,a double-head network decouples classification and regression features,and incorporates a feature refinement module guided by attention mechanism and polarization functions,enabling the construction of task-specific features that facilitate accurate object detection.Experimental results on public remote sensing datasets,including DOTA,HRSC2016,and UCAS_AOD,demonstrate the effectiveness of the proposed algorithm.Compared to the Faster RCNN algorithm,the proposed method achieves accuracy improvements of 8.48%,7.60%,and 3.10%on the three datasets,respectively.The proposed method enables highperformance object detection in remote sensing images.