Detection and Counting Model of Soybean at the Flowering and Podding Stage in the Field Based on Improved YOLOv5

A phenotype survey on soybean flower and pod drop conducted by agricultural experts revealed issues such as poor real-time performance and strong subjectivity. Based on the YOLOv5 detection model, a microscale detection layer is added and the size of the initial anchor box is improved to enhance feature expression ability. The CBAM attention mechanism is introduced in the backbone network to capture the information of direction and position, which helps the model to locate and recognize more accurately. The test results show that the accuracy rate of the soybean flower and pod recognition model reaches 98.4%, and the recall rate reaches 97.4%. Compared with the original network model, the accuracy rate and recall rate increase by 12.8% and 4.1%, respectively. Compared with manual counting, the average accuracy rate of field flower number is 80.32%, and the average accuracy rate of pod number is 82.17%. The research results show that models can effectively replace manual labor to complete the task of field soybean flower and pod identification and counting, and this application will promote the study of the basic laws of flower and pod fall and provide phenotypic investigation techniques.