To achieve rapid and accurate identification of rice diseases in complex field environments， a high-performance lightweight model named YOLO-SEG is proposed， which is improved based on YOLOv11n. Through synergistic optimization， the model integrates a lossless downsampling module （SPDConv） to enhance the feature extraction of tiny disease spots， introduces an efficient multi-scale attention （EMA）module to focus on key features， and adopts lightweight convolution （GSConv） to balance detection accuracy and inference efficiency. On a self-constructed dataset containing six common rice diseases， the mean average precision （mAP@0.5） of the YOLO-SEG model reaches 95.6%， which is 4.4 percentage points higher than that of the baseline model. Meanwhile， the model has only 2.7 MB of parameters and an ultra-high inference speed of 120 FPS. The results show that the YOLO-SEG model achieves a superior balance among detection accuracy， model complexity and inference speed， and provides an efficient and reliable technical support for the real-time disease detection in the field of smart agriculture.