Abstract: The identification and diagnosis of tobacco diseases are the prerequisite and foundation for scientific prevention and control of the diseases. To break the barriers of traditional deep learning methods in segmenting tobacco leaf spots such as weak generalization and sensitivity to noise, four leaf diseases including angular spot, brown spot, wildfire disease and frog eye spot were studied, and the Enhanced Dual Unet(ED-Unet)model based on Unet was constructed to accurately segment these leaf spots by improving the feature extraction module and embedding an attention mechanism. The results showed that: 1)ED-Unet reached 92.75%, 90.94%, 84.93% and 91.81% in indexes of spot CPA, Recall, IoU and F1 score, respectively, with an overall Dice score of 94.67%. 2) The numbers of parameters and floating-point operations and the inference time for a single image of the ED-Unet model were 46.5 M, 233.92 GFLOPs and 65.096 ms, respectively. 3)The accuracy of the ED-Unet model was significantly improved compared with those of Unet, PSP, DeepLab v3 +, FCN, SegNet, UNET ++ and DoubleU-Net, and the ED-Unet model complexity was well controlled, thereby achieving optimal overall performance. This method provides technical supports for the precise segmentation of leaf disease spots of tobacco and other plants.