Abstract: Considering that the regional characteristics of mung bean leaf disease spots with similar disease index are not clearly differentiated, the effect of using fixed scale convolution kernel to detect disease spots with similar size is not very good, so a parallel multi branch multi-scale convolution neural network（PMMS-Net） model is designed. Firstly, the model uses parallel multi branch and multi-scale feature fusion module to obtain rich plaque features. Then the coordinate attention mechanism is used to make the model better locate the lesion area and realize selective emphasis on the region of interest. Finally, full feature extraction module is used to combine the depth separable convolution with the ordinary convolution to achieve the full extraction of features and further optimize the effect of feature extraction. The experimental data set consists of chlorophyll fluorescence images of mung bean leaf spot disease, including five resistance types of mung bean leaf spot disease images. The results show that the method proposed in this paper takes only 0. 8 times more time to train 1 000 iterations on the dataset than AlexNet, but the verification accuracy is 18. 9% higher than AlexNet. The verification accuracy of this model on the dataset is 87. 8%, the average specificity is 96. 92%, and the parameter memory is only 0. 54 MB. The method proposed in this paper is conducive to deploying the model on embedded devices with limited resources, such as mobile terminals, and provides a new method for identification of resistance to mung bean leaf spot.