To solve the unsupervised segmentation problem of clustering diced-potatoes, a watershed and improved Markov random field (MRF) algorithm was proposed. The original image was easily transformed from pixel based to region based by watershed algorithm, which was good for labeling by MRF. At the same time, the ISING model was improved to make the consistent of probability of MRF. Firstly the original image was transformed from pixel based to region-based by watershed algorithm. Secondly the improved MRF was applied to distinguish over segmentation regions from right segmentation regions by fusing the relative height and area of the original segmentation regions. Finally the most compactness adjoining over segmentation regions were connected into bigger ones. Using this algorithm, 95% of the test clusters were correctly segmented in potatoes preparations.