The microstructure is closely related to mechanical strength of the stem, which plays an important role in crop lodging resistance. However, the lack of effective methods in identification and estimation of the parameters severely restricted the related researches. In view of the complexity of wheat stalk cross-section microscopic image data set, ResNet50 and Unet deep learning network were used to build a semantic segmentation model Res-Unet for vascular bundles and background regions. MobileNet and Unet networks was combined to build a cross-section, marrow cavity and background. The semantic segmentation model Mobile-Unet measured the relevant parameters of lodging resistance such as the cross-sectional size of the wheat stem, the size of the pulp cavity and the area of the vascular bundle. For small sample data sets, the trained ResNet50 network weights were applied to the network model of wheat stalk cross-sectional slice images through the shared parameter method of transfer learning in deep learning. The results showed that compared with the previous studies, the key parameters greatly improved in accuracy, and the recognition rate of all parameters exceeded 97%, and the highest was 99.91%. Moreover, it only took 21.6s to detect a single image, which was an average increase of 80.36% over the 110s of existing image processing methods. In addition, the model evaluation accuracy rate, recall rate, F1 value and mean intersection over union (mIoU) index values all reached 90%. In conclusion, the method developed was accurate, real-time and effective, and can serve as one of important techniques for the further studies of crop lodging resistance.