Automatic and accurate estimation of disease severity is critical for disease management and yield loss prediction. Traditional disease severity estimation steps are complicated and inefficient, which makes it challenging to achieve accurate estimation in practical scenarios. A disease severity estimation method was proposed based on mixed dilated convolution and attention mechanism to improve UNet (MA-UNet). Firstly, to solve the problem of different sizes and irregular shapes of lesions, the mixed dilation convolution block (MDCB) was proposed to increase the receptive field and maintain the continuity of lesion information to improve the accuracy of lesion segmentation. Secondly, to overcome the influence of complex background, the attention mechanism (AM) was used to model the correlation between the spatial and channel dimensions. It can obtain the response within each pixel class and the dependency between channels to alleviate the backgrounds influence on network learning. Finally, the ratio of diseased lesion pixels to leaf pixels in the disease segmentation map was calculated to obtain the severity. It was validated based on cucumber downy mildew and powdery mildew images collected under field conditions and compared with fully convolutional network (FCN), SegNet, UNet, PSPNet, FPN, and DeepLabV3+. MA-UNet can meet the segmentation requirements of leaves and lesions in complex environments, with a mean intersection over union 84.97% and a value of frequency weighted intersection over union value of 93.95%. Moreover, it can accurately estimate the severity of cucumber leaf diseases, the correlation coefficient was 0.9654, and the RMSE was 1.0837%. The results showed that MA-UNet outperformed the comparison methods in refining lesion segmentation and accurately estimating disease severity. The research result can provide a reference for artificial intelligence to estimate and control disease severity in agriculture rapidly.