Objective: To solve the subjectivity and instability of the traditional Baijiu picking method " liquor-receiving according to liquor hop", and the problem that the existing machine vision hops classification method is difficult to meet the real-time classification. Methods: The lightweight YOLOv5 takes YOLOv5s as the initial model, uses the K-mean clustering anchor box to replace the default anchor box to improve the model detection accuracy and stability, uses the shufflenetv2 network to replace the YOLOv5s backbone network for feature extraction, so as to achieve the purpose of lightweight model, and adds the CBAM attention mechanism to make the model pay more attention to the characteristics of hops. Results: Compared with the initial YOLOv5s model, the memory occupied by the lightweight YOLOv5 model is reduced by 92.5%, the parameters are reduced by 93.7%, the calculation is reduced by 63.4%, the detection accuracy is improved by 2.8%, and the FPS is up to 526. Conclusion: The lightweight YOLOv5 reduces the requirements for hardware configuration and can well realize the real-time detection and classification of hops.