Abstract: Accurate detection of tea diseases is crucial for a high yield and quality of tea, thereby increasing production and minimizing economic losses. However, tea disease detection faces several challenges, such as variations in disease scales and densely occluded disease areas. To tackle these challenges, a novel method for detecting tea diseases called multi-scale guided self-attention network（MSGSN） was introduced, which incorporated multi-scale guided self-attention. The MSGSN method utilized a VGG16-based module for extracting multi-scale features to capture local details like texture and edges in tea disease images across multiple scales, effectively expressing the local multi-scale features. Subsequently, the self-attention module captured global dependencies among pixels in the tea leaf image, enabling effective interaction between global information and the disease image’s local features. Finally, the channel attention mechanism was employed to weight, fuse, and prioritize the multi-scale features, thereby enhancing the model’s ability to characterize the multi-scale features of the disease and improving disease detection accuracy. Experimental results demonstrated the MSGSN method’s superior detection performance in complex backgrounds and varying disease scales, achieving an accuracy rate of 92.15%. This method served as a valuable reference for the intelligent diagnosis of tea diseases. In addition, the method can provide a scientific basis for the prevention and control of tea diseases and help farmers take timely and effective control measures. At the same time, the method can also provide technical support for the development of the tea industry.