Abstract: The phenotypic characteristics of corn ear are closely related with kernels information of corn ear. For example, ear rows, kernels in row, total kernel number and kernel shape are directly determined by kernels distribution in the surface of corn ear. However, the quantitative analysis of kernels based on images of corn ears is a challenging task owing to shape distortion and color difference. In this paper, we presented a novel segmentation method to extract kernels from the image of corn ear, which was effective to overcome the shape distortion and color differences of kernels. The shape distortion of kernels in images manifests that kernel shape was highly sensitive to the imaging angle and orientation of corn ears, i.e. kernels in the different positions of the same image showed completely different shape characteristics, such as area and aspect ratio etc. Thus, the shape information of kernels needed to be recovered from the image of corn ear. Considering the diversities of corn ears, especially variegated ears, common segmentation methods based on color and threshold parameters were difficult to exactly extract all kernels since the threshold values of variegated kernels were located in the large distribution range. Hence it was necessary to find the local threshold for each kernel in the image of corn ear. The proposed method consisted of three main steps: radial distortion correction of corn ear, hierarchical threshold, and multilevel screening of kernels. In the first step, a radial distortion correction method was developed to recover the surface information of corn ear, which unfolded the surface of corn ear along its radical direction according to the three-dimensional shape characteristics, and generated a corrected image in which the edges of corn ear were extended and the shapes of kernels were restored for the subsequent analysis. In the second step, a hierarchical threshold strategy was applied to iteratively segment kernels from the image of corn ear. Following each threshold process, the geometrical properties (area and perimeter etc.) of segmented objects were respectively calculated and further used to evaluate if they can be used to classify valid kernels. By hierarchical thresholds, the adaptive local threshold for each kernel can be detected. Therefore it was useful for elimination of adhesion effect between kernels. In the last step, feature extraction, principal component analysis (PCA) and support vector machine (SVM) were combined to investigate the validness of segmented kernels. For each kernel, 115 shape, color and texture feature descriptors were extracted, and a 19-dimentional vector was generated based on PCA. Then, a SVM model of kernels was trained and tested using the training samples with 2164 images of kernels. This model was used to filter out invalid segmented objects, e.g. abortive, vacant, or bare area. Finally, segmented objects which were determined as valid kernels were collected to build a distribution image of kernels, which represented kernel information of half corn surface and were quite appropriate for calculating the phenotypic characteristics of corn ear and its kernels. Experimental results demonstrated the proposed method has significant advantages than those that already existing in accuracy and robustness for kernel segmentation of various types of corn ears, thus it can be used as fundamental method for automated trait analysis of corn ear. Under the mode of the highest accuracy, the average computation efficiency was 15 second per ear. In the future work, parallel computation and parameters templates techniques are needed to improve the algorithm efficiency.