Abstract: To address the issues of low computational efficiency, significant manual errors, and limited batch processing capability in the analysis of tobacco vein morphology, a deep learning network based on hyperspectral images was developed and named Tobacco Vein U-Net (TVU-Net) for intelligent measurement of leaf morphological parameters. Principal Component Analysis (PCA) was employed to reduce the dimensionality of the hyperspectral images of tobacco leaves, extracting key feature wavelengths and decreasing data redundancy. The TVU-Net integrated Deformable Convolution (DC), Convolutional Block Attention Module (CBAM), and Atrous Spatial Pyramid Pooling (ASPP) modules to achieve fine-grained vein segmentation. The results showed that: 1) Compared with SegNet, U-Net, U-Net++, TransUnet, and FR-UNet, the TVU-Net achieved the highest mean precision, accuracy, mean intersection over union (mIoU), and mDice coefficient. 2) The TVU-Net produced less segmentation noise, preserved the overall vein structure more completely, and captured both the global information of the main vein and the local details of the secondary veins, thus enabling high-precision segmentation of the complex vein structure. 3) On 50 randomly selected tobacco leaf samples, the mean absolute error in main vein length measurement was 2.67 cm, with an average relative error of 4.710%, indicating high measurement accuracy. This study provides a theoretical basis for the automated morphological analysis and parameter quantification of tobacco veins in industrial applications.