<p>Traditional efficient lightweight image classification algorithms generally demonstrate low accuracy in real-time wetland bird recognition tasks due to the environmental complexity and the high similarity among bird species. Moreover, a bird recognition server needs to perform computation-intensive tasks of multi-process parallel inferences, requiring a low inference latency of the bird recognition algorithm. Traditional high-accuracy fine-grained methods cannot meet the demands due to their high computational complexity. In this study, we introduce a scalable two-stage model for real-time wetland bird recognition, which incorporates an object detector and a fine-grained image recognition technique, bilinear pooling, to encode fine-grained features. Additionally, we design a lightweight architecture and propose a bilinear scalable module in the bilinear pooling to trade-off between latency and accuracy. The experimental results show that the proposed method achieves 77.6% and 97.6% accuracy on the CUB and WPB datasets, respectively, which are much higher than MobileNetV3 and ShuffleNetV2, with a low inference latency of only 79.5 ms on CPU. Furthermore, parallel inference experiments in practical environments demonstrate that the proposed method achieves an inference speed of 15.3 FPS, with 12 parallel video streams.</p>