Abstract: The detection conditions of the actual tea garden field scene are relatively complex, which is easy to cause low contrast between the disease area and the background, blurred edge features of the disease spot, and difficult to identify subtle lesions.Meanwhile, traditional detection models generally suffer from inadequate lightweight design, making them ill-suited for deployment conditions of field-edge devices and limiting their overall engineering applicability.Based on the YOLOv8 basic framework, this paper builds a lightweight and high-precision disease detection model that is more suitable for the complex scene of tea garden.Addressing the challenges in tea disease detection, this study developed the RFCAConv-C2F feature enhancement module, which significantly improves the ability to extract and locate subtle lesions under complex conditions.The backbone network was embedded in the CARAFE up-sampling structure, and the edge details and morphological characteristics of the lesion were completely preserved.By employing the LSDECD lightweight decoder to reconstruct the network decoding architecture, detection performance is significantly improved while reducing computational overhead by 6.2%, achieving simultaneous optimization of accuracy and inference efficiency.From the verification results of the data set, the mAP @ 50 index of the improved model on the self-built tea disease data set reached 91.2 %, and the detection accuracy was 91.5 %.Compared to the baseline model, its mAP@50 score on the public dataset improved by 5.7 percentage points.The overall detection performance of the improved model has obvious advantages, which can be well adapted to the actual detection scene of the complex field in the tea garden.