With the aim to give a more accurate computing result of the maize kernel number based on image analysis, a new segmentation method for touching maize kernels was provided. First, contour smoothing and the Euclidean distance transformation were applied to the binary image of the touching kernels. Then the distance map was regarded as topographic relief and the valley lines were detected as the segmentation lines. Considering the boundary pixels of the distance map as water heads, the water will flow forward. They will turn when facing a slope and will go on when facing a valley mouth according to the terrain of the valley. This results in the segmentation. This method will take no influence on single kernels, so does not need a touching estimate by manual input threshold before segmentation. An experiment to 442 connected components of single kernels and multi touching kernels shows an accuracy of 99.7% for touching estimation and 94.9% for touching kernels segmentation.