Abstract:Addressing the core challenge of model severely degraded detection accuracy in infrared object detection due to scarce training data and insufficient annotations, this paper proposes a cross-modal learning-based infrared target detection method based on an improved YOLO11s. The method mitigates the challenge of scarce infrared data by leveraging the abundance of visible-light data. Specifically, it incorporates: A combination of diverse data augmentation strategies to simulate the infrared domain distribution; A lightweight module that reduces model complexity and optimizes deployment on resource-constrained edge devices; Integration of a partial self-attention mechanism after the C2PSA module to optimize visible-light feature repre-sentation, thereby improving the model’s cross-modal generalization capability for the infrared domain. Experimental results demonstrate that the proposed method achieves the mAP of 52.2%, while maintaining a parameter count of 8.8×10^6 on the Drone Vehicle dataset. Compared to existing mainstream object detection algorithms, it exhibits significant advantages in both model lightweight and detection accuracy, providing an efficient solution for infrared target detection under scarce data scenarios.