A Dynamic Detection Method for Phenotyping Pods in a Soybean Population Based on an Improved YOLO-v5 Network

Pod phenotypic traits are closely related to grain yield and quality. phenotype detection in soybean populations natural environments is important breeding, cultivation, field management. For an accurate pod description, a dynamic method proposed based on improved YOLO-v5 network. First, two varieties were taken as research objects. A self-developed three-dimensional color image acquisition vehicle was used obtain RGB depth images of pods the field. Second, red–green–blue (RGB) registered using edge feature point alignment metric accurately distinguish complex environmental backgrounds establish red–green–blue-depth (RGB-D) dataset for model training. Third, pyramid network path aggregation (FPN+PAN) structure channel attention atrous spatial pooling (CA-ASPP) module introduced improve dim small target detection. Finally, quantity compensation established by analyzing influence number individual plants population precision statistically correct predicted quantity. In experimental phase, we analyzed impact different datasets performance models same under test conditions. The results showed that compared with trained dataset, recall RGB-D increased approximately 32% 25%, respectively. Compared YOLO-v5s, 6%, reaching 88.14% 200 population. After compensation, mean relative errors between actual quantities 2% 3% varieties. Thus, can provide rapid massive phenotyping theoretical basis technical knowledge scientific