Abstract: According to the properties of cotton disease image, we propose a segmentation method under the condition of complex background for improvement on segmentation precision of cotton mite disease image. The proposed algorithm is comprised of three main steps. First, we extract the analogous disease spots (disease spots and stems with red color) from green plants by Excess green feature 2G-R-B, then some different gray-scale images would be obtained. Second, the targets would be extracted from the background by thresholding according to the double-peak feature presented in the gray Histogram of the gray-scale images. After this procedure, we can get the binary image that contained only analogous disease spots. According to the further amplification researches on the binary images of the large quantities of cotton mite disease image samples, we can find that the segmented images are constituted by a large number of independent 8-connected region, and the connected region of stems are larger than cotton mite disease spots’. On the basis of this feature, the image can be grouped into two categories in order to remove the stems from the whole analogous disease spots in the binary image: the one is small connected region composed by cotton disease spots; and the other is large connected region composed by cotton stems. Finally, compare the disease spots’ area with stems’ and then segmented binary images by using area thresholding, By observing the segmentation results in different thresholding values, select a optimal one to eliminate stem regions that larger than the value. On the contrary, the cotton mite disease area that smaller than the value will be remained. The experiment results show that this algorithm is of effective in segmenting cotton disease spot, and the correctness rate of the algorithm can reach 98.1%. At last, In order to test the validity and generality of this proposed method, 30 color images of cotton mite-disease are picked out to segment by the proposed algorithm, a split plot with repeated measures in the error extraction rate was used as the experimental design. From the statistics, we can know that the average error extraction rate of 30 color images of cotton mite-disease is 2.17%, the average correct extraction rate could reach 97.83%. This proposed algorithm combined 2G-R-B、single threshold and area thresholding to segment the disease spots from image with complex background plays well. It can lay a foundation for automatic identification of cotton mite disease.