A High-Performance Thermal Infrared Object Detection Framework with Centralized Regulation

Jinke Li, Yue Wu, Xiaoyan YangarXiv 2025 · DOI · arXiv

Abstract

Thermal Infrared (TIR) imaging is essential for applications in nighttime surveillance, autonomous driving, and search-and-rescue. However, TIR images typically suffer from low contrast and a lack of fine-grained textures, posing significant challenges for standard object detection models. This paper presents CRT-YOLO, a novel framework designed to enhance TIR object detection by addressing feature fusion gaps and computational efficiency. CRT-YOLO introduces Efficient Multi-Scale Attention (EMA) to capture long-range dependencies and a Centralized Feature Pyramid (CFP) to regulate feature maps globally. Evaluated on the FLIR and LLVIP datasets, CRT-YOLO achieves state-of-the-art performance, notably reaching 79.6% mAP50 on FLIR, significantly surpassing the YOLOv8 baseline.

Introduction

Object detection in the Thermal Infrared (TIR) spectrum is critical for operating in environments where visible light is insufficient. Despite the success of models like YOLOv8 in the RGB domain, their application to TIR data is hindered by three primary issues: low contrast and detail, ineffective feature fusion between local and global contexts, and high computational overhead of complex attention mechanisms. Traditional Feature Pyramid Networks (FPNs) often fail to maintain a balance between fine details and the overall heat signature of a scene. To overcome these limitations, we propose CRT-YOLO, which utilizes "centralized regulation" to harmonize local precision with global scene understanding.

Proposed Method

CRT-YOLO builds upon the YOLOv8 architecture, incorporating two key structural innovations: Efficient Multi-Scale Attention (EMA) and a Centralized Feature Pyramid (CFP).

The CRT-YOLO framework overview showing the integration of EMA modules between the backbone and the CFP Neck to process multi-scale features.

Efficient Multi-Scale Attention (EMA)

To avoid the heavy computational cost of Transformers, CRT-YOLO integrates EMA modules. These modules partition feature maps into groups and process them through parallel routes using $1 \times 1$ and $3 \times 3$ convolutions. This design enables the model to capture long-range dependencies while preserving local spatial information without significantly increasing the parameter count.

Centralized Feature Pyramid (CFP)

The core of the framework is the CFP, which implements a "Centralized Regulation" strategy. Unlike standard pyramids that pass information sequentially, the CFP employs:

  • Explicit Visual Center (EVC): This component combines a lightweight Multi-Layer Perceptron (MLP) for global dependencies with a Learnable Visual Center (LVC) to aggregate local features.
  • Global Centralized Regulation (GCR): GCR utilizes abstract features from the deepest network layers to regulate shallower layers simultaneously, preventing information dilution and ensuring global awareness throughout the network.
Detailed architecture of the Centralized Feature Pyramid (CFP) highlighting the Explicit Visual Center (EVC) and the Global Centralized Regulation (GCR) mechanism.

Experimental Results

The model was rigorously tested on two benchmark TIR datasets: FLIR (traffic scenes) and LLVIP (low-light pedestrian detection).

  • FLIR Dataset: CRT-YOLO achieved 79.6% mAP50, outperforming YOLOv8 (76.3%), Faster R-CNN (74.4%), and SSD (65.5%).
  • LLVIP Dataset: The model reached 95.1% mAP50, surpassing the YOLOv8 baseline of 93.9%.

Ablation studies highlighted the importance of the EVC; its removal led to a significant performance drop (from 79.6% to 75.3% on FLIR), confirming that capturing the visual center is vital for thermal image interpretation. Qualitatively, CRT-YOLO demonstrated a superior ability to recover small objects, such as distant pedestrians and bicycles, which were often missed by standard models.

Conclusion

CRT-YOLO provides a robust and efficient solution for TIR object detection. By replacing heavy Transformers with lightweight MLPs and implementing centralized regulation, the framework achieves high accuracy with manageable complexity. Future research will focus on optimizing the model for real-time deployment on edge devices and exploring cross-modal fusion techniques to leverage both RGB and TIR data.

Summary updated

BibTeX
@misc{li2025high,
  author = {Jinke Li and Yue Wu and Xiaoyan Yang},
  title = {{A High-Performance Thermal Infrared Object Detection Framework with Centralized Regulation}},
  year = {2025},
  eprint = {2505.10825},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.10825},
}
Figure