Abstract
Visible-infrared object detection for remote sensing images plays an important role in the unmanned aerial vehicle (UAV) around-the-clock application. Most of the existing work focuses on designing complex network architectures to fuse complementary features, while few methods consider computational complexity and susceptibility against modality attacks, limiting the deployment of state-of-the-art frameworks. In this paper, we present a low-rank multimodal object detection approach with frequency filtering experts, called LF-MDet, which is based on the advanced DINO (detection transformer with improved denoising anchor boxes) framework. This approach achieves more accurate detection with fewer computational resources. In particular, when the specific modality is attacked or missing, our method still maintains higher robustness towards such pervasive perturbations. Specifically, we propose a low-rank enhancement technology and a dynamic illumination-aware mask module to enable a single backbone network in the form of batch formulation to unbiasedly and compatibly extract multimodal features. Furthermore, we design a lightweight frequency expert encoder from the frequency domain perspective to efficiently fuse complementary features by filtering out amplitude noise components and mixing feature tokens. Extensive experiments are conducted on the multimodal remote sensing object detection datasets, VEDAI and DroneVehicle. The results demonstrate the superiority of the proposed approach over advanced multimodal remote sensing object detectors. Compared to the baseline method, our LF-MDet effectively reduces the floating point of operations (FLOPs) by approximately 65% while improving detection accuracy. The code is available at https://github.com/cq100/LF-MDet.
Funding Information
  • National Natural Science Foundation of China (61671219)

This publication has 53 references indexed in Scilit: