In response to the irregular morphology of Allium chinense, which led to the problems such as distorted feature representation and difficulty in extracting external contours during detection, a multi-scale keypoint recognition and classification method based on deep learning was proposed. By identifying the apex and the two endpoints of the transverse diameter of Allium chinense, its external contour features were extracted. Simultaneously, a decision tree algorithm was employed to classify the size and shape characteristics of Allium chinense. Initially, using YOLO 11n-Pose as the baseline model, a BMR-YOLO 11n-Pose model was constructed by introducing a bi-level routing attention (BRA) mechanism in the neck, a feature enhancement layer (MobileNet Variants), and a reparameterized convolution with channel shuffle (RCS). Compared with the baseline model, improvements of 0.9 and 1.9 percentage points in key point recognition accuracy (Pose-P) and bounding box recognition accuracy (Box-P) in the proposed model were achieved respectively, with an average precision (mAP0.5-0.95) increase of 1.9 percentage points. Furthermore, based on the contour features, a decision tree algorithm was applied to classify the size of Allium chinense, the classification model accuracies were 92.87% and 84.72% respectively, representing improvements of 11.19 and 8.39 percentage points over the original baseline model, effectively enhancing the classification accuracy of Allium chinense appearance characteristics. The research result can provide theoretical references and technical support for application scenarios such as Allium chinense pose recognition and classification.