The railway catenary system is highly susceptible to environmental factors due to its long-term outdoor exposure. Foreign objects such as bird nests, that get trapped in the catenary can easily cause anomalies in the pantograph-catenary system. However, due to issues like the small pixel ratio and difficult feature extraction of bird nests and branches, current railway foreign object detection systems still face limitations in accuracy and efficiency when detecting small targets like bird nests. To address these issues, this paper proposes a lightweight single-stage object detection model called SFPA-NNet. By designing a shallow fusion feature pyramid aggregation (SFPA), it enhances the shallow feature information to improve the resolution and feature representation of small targets. Meanwhile, it introduces a loss function based on normalized Wasserstein distance to solve the problem of traditional IoU measurement being sensitive to the scale of small targets. Experimental results show that SFPA-NNet achieves an mAP of 94.5% on the self-built dataset while maintaining a high inference speed of 23.8 frames per second. This meets the requirements for lightweight and real-time detection devices on railway vehicles, providing an efficient and accurate solution for bird nests detection in railway catenaries.