A Cross-Modality Feature Adaptive Interaction Approach for RGB-Infrared Object Detection in Aerial Imagery

초록

Object detection in aerial imagery, particularly from unmanned aerial vehicles (UAVs) and remote sensing platforms, is crucial but faces significant challenges such as modality misalignment, feature fusion degradation, and high computational complexity. To address these issues, this article introduces cross-modality feature adaptive detection (CMFADet), a novel framework for robust RGB-infrared (IR) object detection across diverse aerial scenarios. CMFADet improves feature learning through its innovative spatial-frequency feature enhancement module (SFEM) and IR adaptive feature aggregation block (IR-AFAB). It also integrates a channel interaction fusion (CIF) module for dynamic weight allocation, ensuring truly complementary information integration and avoiding mutual interference. This allocation is governed by the specific characteristics of the target and the inherent strengths of each modality. Detection accuracy is further refined via an adaptive task-aware alignment head (ATAH) that learns the joint features. Extensive experiments on the DroneVehicle, vehicle detection in aerial imagery (VEDAI), and OGSOD-1.0 datasets demonstrate CMFADet's superior performance, consistently surpassing state-of-the-art algorithms, and effectively addressing the aforementioned challenges. The source code for this work is publicly available at https://github.com/Yooyoo95/CMFADet

키워드