Dual Information Complementary Fusion Network (DICFNet) for Multispectral Pedestrian Detection

In the phenimenon of low-light and night scenarios, single visible-image pedestrian detection suffers from weak textures and poor target visibility, while thermal images resist lighting variations and offer supplementary localization clues. Fusing RGB and thermal data is critical for robust detection under complex illumination, yet naive feature concatenation or summation brings redundant signals and background noise, failing to fully utilize cross-modal complementarity. This work proposes Dual Information Complementary Fusion Network (DICFNet) for multispectral pedestrian detection. It employs a two-branch feature extraction structure equipped with a Dual Information Complementary Fusion (DICF) module to realize cross-modal feature interaction. Inside DICF, intra-modal channel attention highlights valuable features of each modality, and inter-modal differential information supplements unique modal characteristics. The two processed signals are fused with original feature maps to generate highly distinguishable representations. Moreover, a multi-label enhancement learning scheme is adopted to deliver targeted supervision for diverse RGB-thermal matching states, thus enabling the model to handle misaligned or incomplete cross-modal inputs. Evaluations on KAIST and CVC-14 datasets show that DICFNet outperforms mainstream competitors. On the KAIST dataset, DICFNet achieves the lowest miss rate of 7.26% on the reasonable subset and 5.16% at night, significantly outperforming MBNet (8.13%/7.86%) and MLPD (7.58%/6.95%). It also demonstrates exceptional robustness under heavy occlusion (Heavy: 53.74% vs. 59.14% for MBNet and 60.28% for MLPD) and far distance (Far: 52.36% vs. 55.99%/57.45%). On the CVC-14 dataset, DICFNet attains the best overall and nighttime performance with miss rates of 20.94% and 13.41%, surpassing MBNet (21.10%/13.50%) and AR-CNN (22.10%/18.10%). Ablation studies confirm that both DICF module and multi-label learning boost accuracy, proving that the combination of intra-modal weight optimization and inter-modal complement extraction benefits multispectral pedestrian detection.

Authors

Institutions

Publication Details

Journal
Machines
Published
2026-11-16
DOI
https://doi.org/10.3390/machines14111354
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Dual Information Complementary Fusion Network (DICFNet) for Multispectral Pedestrian Detection

Yanting Wang, Yicheng Li
Machines
Advanced Neural Network Applications
article

Dual Information Complementary Fusion Network (DICFNet) for Multispectral Pedestrian Detection

Yanting Wang, Yicheng Li
article en

Abstract

In the phenimenon of low-light and night scenarios, single visible-image pedestrian detection suffers from weak textures and poor target visibility, while thermal images resist lighting variations and offer supplementary localization clues. Fusing RGB and thermal data is critical for robust detection under complex illumination, yet naive feature concatenation or summation brings redundant signals and background noise, failing to fully utilize cross-modal complementarity. This work proposes Dual Information Complementary Fusion Network (DICFNet) for multispectral pedestrian detection. It employs a two-branch feature extraction structure equipped with a Dual Information Complementary Fusion (DICF) module to realize cross-modal feature interaction. Inside DICF, intra-modal channel attention highlights valuable features of each modality, and inter-modal differential information supplements unique modal characteristics. The two processed signals are fused with original feature maps to generate highly distinguishable representations. Moreover, a multi-label enhancement learning scheme is adopted to deliver targeted supervision for diverse RGB-thermal matching states, thus enabling the model to handle misaligned or incomplete cross-modal inputs. Evaluations on KAIST and CVC-14 datasets show that DICFNet outperforms mainstream competitors. On the KAIST dataset, DICFNet achieves the lowest miss rate of 7.26% on the reasonable subset and 5.16% at night, significantly outperforming MBNet (8.13%/7.86%) and MLPD (7.58%/6.95%). It also demonstrates exceptional robustness under heavy occlusion (Heavy: 53.74% vs. 59.14% for MBNet and 60.28% for MLPD) and far distance (Far: 52.36% vs. 55.99%/57.45%). On the CVC-14 dataset, DICFNet attains the best overall and nighttime performance with miss rates of 20.94% and 13.41%, surpassing MBNet (21.10%/13.50%) and AR-CNN (22.10%/18.10%). Ablation studies confirm that both DICF module and multi-label learning boost accuracy, proving that the combination of intra-modal weight optimization and inter-modal complement extraction benefits multispectral pedestrian detection.

MachinesVol. 14(11)
Jiangsu University (CN)
National Natural Science Foundation of China
Sustainable cities and communities
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon