The composition and fine structure of corn kernel are directly related to yield and quality of maize. An automatic measurement method for threedimensional structures of corn kernels based on CT images was proposed, which could quickly extract and analyze the components and structural traits of corn kernel, and the differences of traits among different maize varieties were evaluated. Firstly, CT images of batch corn kernels were obtained by Micro-CT, and single kernel was accurately segmented by using Watershed algorithm. Furthermore, the improved RAUNet-3D network based on attention mechanism was designed to accurately extract kernel embryos. Finally, an automatic phenotype pipeline was established to calculate 23 traits related to corn kernel, embryo, endosperm and cavity, which were used for traits analysis and variety identification of maize kernel. A total of 120 kernels of four corn varieties (Denghai 605, Jingke 968, Syngenta 408 and Nonghua 5) were selected for performance verification. The experimental results showed that the data acquirement efficiency was improved to about 1min per kernel, and the efficiency of algorithm pipeline was about 10s per kernel, and the segmentation accuracy of kernel embryos reached 93.4%. The determinant coefficients of kernel length, width and thickness were 0.902, 0.926 and 0.904, respectively. The proposed method could quickly and non-destructively measure the three-dimensional structure of corn kernel and its components, and the extracted traits could represent the phenotypic differences among different corn varieties, which laid a foundation for large-scale three-dimensional phenotypic measurement and identification of corn kernels.