Abstract: Visual features of different disease types can often exhibit high similarity in the detection tasks of the tomato leaf diseases. Particularly, some challenges still remained for the small-target diseases. Spatial areas are limited in the indistinct features to accurately identify. Deep learning models can be a great potential for plant disease detection with their powerful feature extraction. However, it is typically required for a substantial amount of the high-quality training samples. It is still lacking in the large-scale and well-annotated disease samples in practice. In this study, an improved version of the YOLO v8a model was proposed to detect tomato leaf diseases using super-resolution enhancement. A Super-Resolution Generative Adversarial Network (SRGAN) was employed to enhance the images of tomato leaves. The super-resolution reconstruction effectively amplified the subtle features of the small target diseases, thereby improving the detection accuracy. Furthermore, the detection model named G-YOLO v8 was developed to detect similar disease features and the indistinct characteristics of small target diseases. A small target detection layer (Upsample-Concat-C2f) was integrated with a Triplet Attention module. The small-target detection layer was incorporated with upsampling, feature concatenation, and integration operations. The representational capacity was then achieved for the small-target features. A triplet attention mechanism was incorporated to capture the cross-dimensional dependencies among three feature branches. The network was significantly improved to identify the subtle differences among various disease types. The public datasets were pre-trained to obtain the optimal initial parameters. The insufficient disease samples were addressed in the practical scenarios. Additionally, transfer learning was employed to enhance the accuracy of the detection. Tomato diseases were identified within natural environmental backgrounds. Comparative experiments were conducted on both public and self-constructed datasets. Three types of diseases included the early blight, late blight, and gray mold. Visualization results demonstrated that the disease samples from the public dataset were significantly enhanced after SRGAN network processing. Among them, the small-target diseases were effectively magnified from the public dataset. The YOLO v8 network successfully identified the previously undetectable features of the small-target disease. The quantitative results show that the average accuracy of the YOLOv8 network in the detection of early blight after super-resolution enhancement reaches 95.4%, which is 4 percentage points higher than that of the unprocessed public dataset of early blight. The proposed G-YOLOv8 network has an average accuracy of 97.8% after transfer learning, which is 3.2 percentage points higher than that of the original YOLOv8 network. This finding can offer a viable solution and technical support to the visual detection of the tomato leaves, particularly for the multiple types and small target diseases.