基于多模态交互引导网络的遥感图像密集车辆检测

Remote Sensing Image Dense Target Detection Based on Multimodal Interactive-guided Network

遥感图像目标检测是大规模拥挤城市场景的风险评估和救援的技术基础,面临着光照干扰、遮挡、密集、拥挤等挑战。现有工作主要基于可见光单模态的遥感数据,对密集目标的特征表达及可分离能力受限。提出了一种面向遥感车辆检测的多模态交互引导网络,通过构建双流网络及多模态交互引导学习机制,显著提升密集车辆目标的可分离度,解决由于类间距离小、类内距离大,导致密集车辆的检测性能差的难题。在Potsdam公开数据集和DLR音乐节数据上的实验表明了该算法的鲁棒性和有效性。

Object detection in remote sensing images is the technical basis for risk assessment and rescue of large-scale crowded urban scenes.It faces such challenges as light interference,occlusion,density crowdedness and others.The existing work is mainly based on visible light monomode remote sensing data with a limited feature representation of dense targets and object separability.A multimodal interaction-guided network for remote sensing vehicle detection is proposed.By building a dual-flow network and a multimodal interaction-guided learning mechanism,the separability of dense vehicle targets can be significantly improved and the issue of poor detection performance of dense vehicles caused by small inter-class distance and large intra-class distance are addressed.The experiments on the Potsdam public datasets and DLR music festival data demonstrate the robustness and effectiveness of the proposed algorithm.