Segmentation of gradient images in the traditional watershed algorithm usually has no markers, which is likely to cause excessive segmentation. This paper presents a watershed algorithm based on the label. First, the gradient images are obtained through the reconstruction of gradient by using Sobel operator and, at the same time, the perimeter, area and morphology factors of the region are computed and labeled. Then, the initial image of segmentation is acquired by using the method of distance function icon for determining the seeds and the technique of isoline tracking. Finally, the last result of segmentation is obtained by using an improved method of regional merging. The experimental result shows the effectiveness of the method.