遥感
计算机科学
衰减
解耦(概率)
对象(语法)
遥感应用
雷达跟踪器
雷达成像
合成孔径雷达
目标检测
杂乱
作者
Xiaomin Liu,Haojie Yin,Haoyu Wang
标识
DOI:10.1109/tgrs.2026.3742556
摘要
Tiny objects in multi-modal unmanned aerial vehicle (UAV) remote sensing images are highly susceptible to being overwhelmed by the high-frequency clutter of complex terrains. Existing visible-infrared fusion methods are predominantly confined to the spatial domain, where direct interaction between large-area infrared thermal responses and fine visible structures may attenuate the structural cues required for tiny-object detection. To address structural degradation during multimodal feature interaction, this paper proposes a frequency-domain amplitude–phase decoupling network (FADN) for UAV-based RGB–IR tiny-object detection. FADN introduces frequency-domain interaction as a complementary representation to spatial-domain feature learning. Specifically, an uncertainty-aware phase–amplitude cross-modulation module employs the visible phase as the dominant structural anchor and incorporates complementary infrared amplitude information together with a bounded phase residual according to the estimated infrared confidence. A radial frequency-band recalibration mechanism is further introduced to adaptively model the responses of different radial frequency bands. In addition, a frequency-spatial synergistic interaction module is constructed, in which frequency-domain responses guide adaptive spatial feature sampling, while spatial contextual information modulates the frequency-domain representation. Finally, a frequency-domain structural constraint loss function is designed. By combining RGB-phase consistency, infrared-confidence regularization, and a cross-domain diversity term based on absolute cosine similarity, it regulates multimodal feature learning and reduces representation redundancy. Extensive experiments on the large-scale cross-modal remote sensing datasets DroneVehicle and VTUAV-det demonstrate that FADN effectively alleviates the high-frequency feature decay problem and significantly improves the detection performance of tiny objects in complex remote sensing scenarios.
科研通智能强力驱动
Strongly Powered by AbleSci AI