Aiming to solve the problem of low efficiency and strong subjectivity in human visual identification of diseased tomato plants in natural environments, a tomato yellow leaf curl virus disease grading detection model based on improved YOLO v7 was proposed to detect mild, moderate, and severe diseased plants. The model perception of complex lesion areas was enhanced by introducing DCN modules into the backbone network. Furthermore, some ordinary convolutions were replaced by Pconv modules in the main network to more efficiently extract spatial features and reduce redundant calculations and memory access. A SimSPPF module was introduced into the detection head to reduce the amount of computation, increase the receptive field, and enhance feature extraction capabilities. After testing, the average precision of the improved YOLO v7 model for mildly infected, moderately infected and severely infected tomato plants was 97.5%, 92.1% and 93.6%, respectively. The improved YOLO v7 model showed a mean average precision of 95.0%, which was 0.8 percentage points higher than that of the original model. The number of parameters was reduced by 8.2×105, the number of floatingpoint operations was reduced by 2.7×1010, and the model memory usage was reduced by 15.7MB. The size of the model was reduced while ensuring detection accuracy. Compared with Faster R-CNN, YOLOX, YOLO v5l, and YOLO v8m models, the mean average precision was increased by 11.2, 5.7, 1.4, and 8.7 percentage points, respectively. The experimental results demonstrated that the model can achieve graded detection and identification of tomato yellow leaf curl virus disease, providing support for the intelligentization of tomato cultivation.