In order to better segment the target in the image, the article improves the method of maximum inter-class variance. According to many numbers of segmentation, we used repeatedly the maximum between-cluster variance to segment the sub-image. There were many factors, such as variance, standard deviation, proportions of the area extracted in the original image, and so on. With the factors, we used Bayesian theory to predict and extract the target in the image. This method not only has a better effect on the images with obvious differences between the target and the background, but also has a better effect on the image in which the target gray level does not belong to the maximum or the minimal class. The algorithm is proved effective by experiments.