CFGPNet: Cross-Attention-Based Fused Gradient Programmed Network Framework for Multispectral Object Detection

Multispectral object detection aims to combine visible and infrared imagery to maintain robust performance in challenging environments like low-light, adverse weather, or complex scenes. However, existing methods often struggle with high computational costs, unstable fusion due to modality distribution gaps, and inefficient gradient flow during training. This paper introduces CFGPNet to address these challenges by balancing detection accuracy with computational efficiency.

CFGPNet employs a dual-stream architecture that processes visible and thermal inputs through an improved GELAN backbone, which utilizes RepViT-style re-parameterized blocks to maintain efficiency. The framework introduces three key innovations:

• CrossCEA (Cross Computation Efficient Attention): A module designed to enhance cross-modal interaction while reducing redundant information transfer between the two branches.
• ASAF (Attention Selection and Aggregation Fusion): A network that combines dense feature aggregation with selective attention to produce compact, discriminative fused representations.
• Programmable-Gradient Auxiliary Branch: An auxiliary structure integrated into the network to improve gradient delivery and optimization quality during training.

The authors evaluated CFGPNet on five public multispectral benchmarks: FLIR, M3FD, LLVIP, VEDAI, and MFAD. The framework consistently achieved strong performance across diverse scenes and object scales. For example, it reached 80.7% mAP50 on the FLIR dataset and 97.8% mAP50 on the LLVIP dataset. The results demonstrate that CFGPNet provides a practical, effective solution for multispectral detection, offering a favorable trade-off between accuracy and computational cost across different model scales.

By addressing the core bottlenecks of multispectral detection—specifically the instability of cross-modal fusion and the computational overhead of standard attention mechanisms—CFGPNet provides a scalable architecture suitable for real-world deployment. Its ability to maintain high accuracy while remaining efficient makes it a valuable contribution for applications requiring robust perception in varying environmental conditions.