DF-Net: A Dual-Modal Detection Network Based on Decoupled Deformable Dilated Convolution and Progressive Coordinated Fusion
IEEE Sensors Journal · 15.06.2026 · 26 (12) · 18811-18823
DOI: 10.1109/jsen.2026.3690667 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Sensors Journal, erschienen am 15.06.2026. Verfasst von Yang Liu und 3 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Natural Science Foundation of Shandong Province, Key Technology Research and Development Program of Shandong Province, Basic Research Project of Yantai Science and Technology Innovation Development Plan.
Abstract (Originalsprache)
Effectively extracting and fusing features between visible and infrared modalities is critical for enhancing object detection performance in complex environments. However, prevailing dual-modal detection models often inherit feature extraction architectures primarily optimized for visible-light textures. While standard convolutions excel at capturing fine-grained RGB details, they may not fully accommodate the holistic morphology and thermal radiation variations characteristic of infrared imagery. Such a homogeneous treatment of different modalities constrains the model’s ability to extract deep complementary features to a certain extent. Furthermore, in the feature fusion stage, existing interaction mechanisms tend to overlook the balance between local detail supplementation and global semantic correlation, thereby limiting fusion performance. To address these challenges, we propose DF-Net, a dualmodal detection network based on decoupled deformable dilated convolution (DDConv) and progressive coordinated fusion. First, to mitigate modality-mismatch in feature extraction, we design the decoupled DDConv for the infrared branch. DDConv utilizes orthogonal 1-D dilated convolutions to match the anisotropic shapes of IR targets (e.g., pedestrians and vehicles) and incorporates a deformable mechanism to adaptively capture thermal gradient edges. Second, we construct progressive fusion module with coordinated local–global (Fusion-LG) attention. This module progressively decomposes the fusion task into three stages: intermodal interaction, intramodal coordination, and dual-modal aggregation. Fusion-LG module leverages cross-attention and a Transformer to capture essential local features and contextual information. Building on this, an adaptive weighting mechanism coordinates the contributions of the original features, local details, and global semantics within each modality. In the aggregation stage, the two enhanced unimodal representations are integrated to output the final multimodal fused features. Third, experiments on three multimodal datasets—M3FD, LLVIP, and VEDAI—demonstrate that DF-Net achieves mean average precision (mAP50–95) improvements of 2.5%, 2.2%, and 2.7%, respectively, compared with the multimodal YOLO11 baseline.
Förderung
- Natural Science Foundation of Shandong Province
- Key Technology Research and Development Program of Shandong Province
- Basic Research Project of Yantai Science and Technology Innovation Development Plan
Zitieren
APA
Liu, Y., Sun, G., Guo, Q., & Wu, Y. (2026). DF-Net: A Dual-Modal Detection Network Based on Decoupled Deformable Dilated Convolution and Progressive Coordinated Fusion. IEEE Sensors Journal, 26(12), 18811-18823. https://doi.org/10.1109/jsen.2026.3690667
Vancouver
Liu Y, Sun G, Guo Q, Wu Y. DF-Net: A Dual-Modal Detection Network Based on Decoupled Deformable Dilated Convolution and Progressive Coordinated Fusion. IEEE Sensors Journal. 2026;26(12):18811-18823. doi:10.1109/jsen.2026.3690667
BibTeX
@article{Liu2026,
author = {Liu, Yang and Sun, Guidong and Guo, Qiang and Wu, Yuzhou},
title = {DF-Net: A Dual-Modal Detection Network Based on Decoupled Deformable Dilated Convolution and Progressive Coordinated Fusion},
journal = {IEEE Sensors Journal},
year = {2026},
volume = {26},
number = {12},
pages = {18811--18823},
doi = {10.1109/jsen.2026.3690667}
}
EUVIMED