IV-PAFNet: Infrared-visible cross-modal alignment and fusion network for nighttime UAV object detection
- 1School of Computer Science and Technology, Shandong University of Technology, Zibo, Shandong, China.
Abstract
Nighttime unmanned aerial vehicle object detection is challenged by degraded visible-light textures, cluttered infrared thermal responses, and local spatial discrepancies between infrared and visible images. To address these issues, we propose IV-PAFNet, an infrared-visible purification, alignment, and fusion network designed for object detection in nighttime and low-visibility conditions. IV-PAFNet consists of two primary components. The infrared signal purification and spatial alignment module applies learnable high-pass filtering and gated residual fusion to refine infrared responses, followed by visible-guided offset prediction and differentiable resampling to alleviate local cross-modal spatial discrepancies. The multi-scale fusion attention enhancement module performs channel-wise modality recalibration and spatial reliability weighting at multiple feature-pyramid levels, allowing the network to adapt the contribution of each modality to local imaging conditions. In the training-based DroneVehicle and generalization-based Camera-vehicle1 datasets, IV-PAFNet achieves 62.81% [email protected]:0.95 and 51.4% [email protected]:0.95, respectively. Additional controlled experiments investigate the distinct roles of infrared purification, visible-guided alignment, channel recalibration, and spatial reliability modeling, while statistical and computational analyses characterize run-to-run variation and inference cost. The code is available at https://github.com/iv-pafnet/IV-PAFNet.
Related
- DHAFGan: A dense hybrid attention fusion generative adversarial network for infrared and visible image fusion
- Visible/Near infrared spectroscopy in muscle beef analysis: how device choice and sampling affect prediction accuracy
- A comparison of artificial neural networks and partial least squares modelling for the rapid detection of the microbial spoilage of beef fillets based on Fourier transform infrared spectral fingerprints
- An annotated cucumber leaf image dataset for detection of visible damage and abnormal color symptoms in complex cultivation environments
- EMAP-SSN: An Embedding- and Multiple-Alignment-Integrated Sequence Similarity Network Platform for Interactive Exploration of Protein Sequence Space
BibTeX
@article{yuan2026ivpafnet,
title = {{IV-PAFNet: Infrared-visible cross-modal alignment and fusion network for nighttime UAV object detection}},
author = {Cheng Yuan and Shichang Shang and Jinghao Zhang and Yang Zhang and Bin Zhou},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0361401},
url = {https://doi.org/10.1371/journal.pone.0361401},
note = {CC BY},
}
RIS
TY - JOUR
TI - IV-PAFNet: Infrared-visible cross-modal alignment and fusion network for nighttime UAV object detection
AU - Cheng Yuan
AU - Shichang Shang
AU - Jinghao Zhang
AU - Yang Zhang
AU - Bin Zhou
PY - 2026
DA - 2026/10/28
JO - PLOS One
DO - 10.1371/journal.pone.0361401
UR - https://doi.org/10.1371/journal.pone.0361401
AB - Nighttime unmanned aerial vehicle object detection is challenged by degraded visible-light textures, cluttered infrared thermal responses, and local spatial discrepancies between infrared and visible images. To address these issues, we propose IV-PAFNet, an infrared-visible purification, alignment, and fusion network designed for object detection in nighttime and low-visibility conditions. IV-PAFNet consists of two primary components. The infrared signal purification and spatial alignment module applies learnable high-pass filtering and gated residual fusion to refine infrared responses, followed by visible-guided offset prediction and differentiable resampling to alleviate local cross-modal spatial discrepancies. The multi-scale fusion attention enhancement module performs channel-wise modality recalibration and spatial reliability weighting at multiple feature-pyramid levels, allowing the network to adapt the contribution of each modality to local imaging conditions. In the training-based DroneVehicle and generalization-based Camera-vehicle1 datasets, IV-PAFNet achieves 62.81% [email protected]:0.95 and 51.4% [email protected]:0.95, respectively. Additional controlled experiments investigate the distinct roles of infrared purification, visible-guided alignment, channel recalibration, and spatial reliability modeling, while statistical and computational analyses characterize run-to-run variation and inference cost. The code is available at https://github.com/iv-pafnet/IV-PAFNet.
SN - 1932-6203
ER -