In order to solve the problem of inaccurate detection of small spots in rice, a rice leaf disease detection method Rice-YOLOv3 based on the improved YOLOv3 was proposed in this study. First, the K-means++ clustering algorithm was used to compute the new anchor frame size for data matching. Second, the activation function Mish was used to replace the Leaky Relu activation function in the YOLOv3 backbone network with a goal to improve the detection accuracy of the network by use of the smoothing property. And, the CSPNet was combined with the residual module in DarkNet53 to avoid the repetition of the gradient information and increase the learning ability of the neural network to improve the detection accuracy and speed. Finally, the attention mechanism ECA and CBAM modules were introduced at the FPN layer to solve the feature extraction problem at the feature layer stacking and improve the detection ability of small spots. In the training process, the COCO dataset was used to pre-train the network model to get the pre-training weights and improve the training effect. The results showed that the mean average precision mean(mAP) of Rice-YOLOv3 for in the rice leaf disease amounted to 92.94%, of which the mAP values of rice blast, brown spot and leaf blight reached 93.34%, 89.68%, 95.80%, respectively. Compared to the YOLOv3, the mAP of Rice-YOLOv3 detection increased by 6.05 percentage points and the speed was improved by 2.8 frames/s, and the detection ability of small spots of rice blast and brown spot was significantly enhanced including those small spots missed by the original network model. Comparing with the models of Faster-RCNN, YOLOv5, etc., the Rice-YOLOv3 improved the ability of recognizing the similar and tiny diseases as well as the detection speed.