Stem microstructure is closely related to its mechanical properties and affects lodging resistance in crops. But crop microphenotypic parameters are difficult to obtain manually. Therefore, automated measurement methods are urgently needed. The lack of measurement methods for high-throughput vascular bundle parameters seriously restricts the in-depth study. Based on the deep learning architecture, ResNet and Unet network were merged to construct the semantic segmentation model Res-Unet to segment function zones in maize stem cross section. In view of the small area, large number and dense distribution of vascular bundle in maize stem cross section, EfficientDet was used as the basic network architecture. According to the characteristics of small size of vascular bundles, the number of layers of BiFPN was reduced to improve reasoning speed and reduce the occupation of video memory. Mask segmentation branches were added to construct a new network Eiff-BiFPN to segment vascular bundles. The results showed that the DICE of each function zone could reach an average DICE of 8817%, and the vascular bundle segmentation task could reach 88.78% and 72.80% on AP50 and AP50:70, respectively. Therefore, the proposed method was accurate, real-time and available, which can be used for automatic determination of microstructure parameters of maize stem, and a technical basis was established for the study of crop lodging resistance. According to the segmentation results, the cross-sectional size of corn stem, the size of each functional area, the number and area of vascular bundles and other microstructure parameters can be obtained.