Soybean(Glycine max) leaf wormholes seriously affect the quality of crops. However, due to the complex background environment, dense planting and diversified leaf wormhole shapes, traditional manual and machine learning recognition are difficult to meet the requirements in terms of accuracy and speed. In response to this problem, this paper proposes an improved soybean pest identification method. This method is based on the YOLOv5s (You Only Look Once) network, introduces an attention mechanism to improve the recognition ability of wormhole parts, uses the sample transformation method to adapt to the diversity of multi-leaf morphology, and improves the redundant bounding box The elimination mechanism reduces misjudgments and missed judgments. In the experiment, this paper constructed a soybean sample data set as the test data, and compared this method with the traditional deep target recognition method. The average accuracy rate on the test data set is up to 95.24%, and the model storage space is 15.1 MB, the number of frames transmitted per second is 91 f/s. The average accuracy rate is 2.50%, 12.13%, 2.81% higher than Faster R-CNN, YOLO v3, and YOLO v5s respectively. The method proposed in this paper has greatly improved the recognition accuracy and recognition speed, and only requires a small model deployment. The above features make this method more suitable for the practical application of soybean wormhole recognition.