A Multi-Level Fusion-Enhanced YOLOv8 Framework for Multispectral Object Detection

In order to address the issue of insufficient feature complementarity between visible-infrared modalities and low computational efficiency in complex environments, a multi-level fusion-enhanced dual-modal object detection framework is proposed. Based on YOLOv8, we design a heterogeneous dual-stream backbone for modality-specific feature extraction, a Squeeze-and-Excitation Fusion module for dynamic channel enhancement, and a Two-stream Fusion Transformer for long-range dependency modeling. Additionally, a lightweight C2f _Faster module optimises computational efficiency in the infrared branch. Experiments on the LLVIP dataset demonstrate that our method significantly improves detection accuracy and real-time performance by effectively fusing multi-modal complementary information, it achieves an mAP50 of 95.7% (1.0% higher than the baseline dual-stream model) with only 6.42M parameters, providing an efficient solution for robust detection in dynamic open-world scenarios.