To address the challenges of low identification rates for grape leaf diseases with similar symptoms appearances and the difficulty in detecting minor diseases, an improved real-time detection transformer(RT-DETR) network-based method for grape leaf disease detection was proposed. First, the coordinate attention(CA) mechanism was employed to enhance the deformable convolutional network v2(DCNv2) module, resulting in the DCNv2_CA module to strengthen target feature extraction capabilities. This module was integrated into the model’s backbone feature extraction layer to improve the model’s ability to extract deep-level key features of plant diseases. Second, the model incorporated the high-low frequency feature interactions(HiLo) attention mechanism within its feature interaction module, which enabled the model to simultaneously focus on both high-frequency and low-frequency feature information, thereby enhancing its ability to detect minor grape diseases. Finally, the cross-layer fusion network of the model was restructured by using the gather-and-distribute mechanism to enable more comprehensive integration of information across different layers, thereby further enhancing the model’s performance in identifying diseases with similar symptoms appearances. The results indicated that the improved RT-DETR model achieved disease detection accuracy, recall, and mean average precision of 90.8%, 89.5%, and 93.4%, respectively. Compared to the initial model, these metrics improved by 5.4, 3.9 and 5.6 percentage points, respectively, and also demonstrated significant advantages over other models. The improved RT-DETR model could accurately detect grape leaf diseases.