Abstract: The watershed transform with gradually changing threshold was proposed for image segmentation of corn count using image processing method. First, binary image was processed with Euclidean distance transform, and regions which values higher than the initial threshold in this image were merged, then the image was segmented by watershed transform. To avoid over-segmentation, single seed regions were cut from source image and pasted in the result image. After that if source image was not empty, was threshold increaseed and the above process was repeated. Finally, the corn seeds amount of the result image was counted. In this paper, 50 images with about 500 corn seeds were tested. The correction rate of segmentation was 97.7%. The method can resolve the problem of separating touching corn effectively and is applicable in corn kernels counting.