Abstract: In order to realize the automatic, universal and accurate image segmentation of the cotton leaves under the natural lighting condition, a co-segmentation algorithm was put forward based on the optimization model of Markov random field. We made use of a co-saliency detection algorithm, which was unsupervised, to generate the co-saliency map for each blade in an set including multiple images of cotton seedling leaves. These maps were used to construct an interior energy function over Markov random field since the co-saliency was able to highlight the common objects（the leaves）. Furthermore, we modeled the dissimilarity between foregrounds of each blade and the common objects in the set by Gaussian mixture models, which were employed to form a global energy item. Finally, we iteratively minimized the whole energy function by means of graph cuts to segment the leaves from the backgrounds in all images. We took 600 images under various imaging conditions, and evaluated the segmentation performance of the proposed algorithm. Experimental results showed that, for the images taken in sunny day, cloudy day and after raining, the average accuracies of segmentation of the algorithm reached 84.8%, 87.7% and 91.6% respectively, improved by 10.7%, 3% and 10% respectively compared with famous Grabcut.