Abstract: Abstract: Although the performance of the deep network-based image recognition model of vegetable leaf disease is excellent, it is still difficult to deploy to the edge computing equipment, embedded equipment, mobile equipment, and other areas of agricultural Internet of Things (IoT) due to the problems of huge parameters, long training time, high storage cost, and calculation cost. Therefore, how to reduce the size of the model while ensuring the accuracy of model recognition has become a hot issue. Based on the study of the advantages and disadvantages of ResNet18, this study proposed an improved Multi-scale ResNet lightweight disease recognition model. The Multi-scale ResNet had made improvements in network structure design, multi-scale feature extraction, feature mapping dimensionality reduction, and complexity reduction. In order to test the effectiveness of the network, this study used two datasets. Dataset 1 used 15 vegetable diseases in the public dataset of the Plantvillage and the AI Challenge2018, and dataset 2 was self-collecting 7 disease images. Each image was resized to 224 × 224 using bilinear interpolation. In order to prevent overfitting due to too few data, the original dataset was enhanced through translation, scaling and rotation, random clipping and scaling, random brightness contrast enhancement, random gamma noise, and vertical flip. The original dataset was expanded to 134 232, and the training set, verification set, and test set were divided according to the ratio of 7∶2∶1. The experimental scheme included model structure design and comparison with other models. In the model structure design experiment, five options were included convolution kernel design, channel design, residual layer connection design, grouping design, and residua block number design. The effectiveness of network design mode was verified by changing one design method by fixing other design methods. Finally, the network structure of Multi-scale ResNet was determined, including the network structure of using large convolution kernel and group convolution, two channels, two-layer connection mode, and two residua blocks. The accuracy of Multi-scale ResNet on dataset 1 and dataset 2 reached 95.95% and 93.05% respectively. Compared with other models (AlexNet, VGG16, ResNet50, SqueezeNet), the improved Multi-scale ResNet has the least number of parameters and the smallest model volume. And the accuracy of Multi-scale ResNet was 3.72% lower than the original ResNet18, but the training parameters of the model were reduced about 93% and the overall size of the model was reduced about 35%. The minimum size of Multi-scale ResNet was 68.75 MB, and the accuracy was only about 1.5% lower than ResNet50. Moreover, the loss value of the model decreases smoothly, and there was no similar oscillation phenomenon of the original ResNet18 and ResNet50. Experiments showed that the Multi-scale ResNet has the characteristics of small size and high accuracy. It made the vegetable leaf disease identification model had the ability to deploy and run in the scene of limited hardware. It overcame the shortcomings of the traditional depth model which was not suitable for edge deployment because of its large parameters and calculation. It could meet the urgent need of realizing long-term, large-scale, and low-cost automatic identification of vegetable diseases under the condition of resource constraints.