Tea plant diseases and pests are the main factors affecting tea production and quality. The accurate detection  of tea diseases and pests is one of the current hot issues in China. Aiming at the problem that the traditional target detection network model is difficult to deploy in industry due to the large number of parameters and low accuracy, the phenotype image dataset of tea diseases and pests is established, the network model is lightened, and the multi-knowledge transfer training model based on knowledge distillation is optimized. A YOLOv5 target detection network model based on the visual attention module （CSA） is constructed, and the tea disease and pest detection method is optimized. The experimental results show that the YOLOv5 target detection model with the visual attention module （CSA） constructed in this paper, compared with the YOLOv5 network model, and the YOLOv5 network model with the traditional attention modules SE and CBAM, respectively, improves the average accuracy by 3.1%, 1.1%, and 1%. Compared with the pre-distilled student model, the model constructed in this paper achieves a maximum accuracy improvement of 4.1%. Compared with the teacher model, the model capacity is reduced by 5.4 MB, and the single-frame image inference time is reduced by 35%. The network model designed in this paper reduces the computational overhead without sacrificing accuracy and can provide an implantation possibility for resource-limited edge computing systems in the field of agricultural informatization.