Abstract: In order to improve the wide adaptability and the real-time processing property of image threshold segmentation algorithm, a 2D maximum between-cluster variance image segmentation algorithm was brought forward. On the base of 2D maximum between-cluster variance algorithm, the impact of the size of the neighborhood template on the best threshold value was studied, and not only the gray level information of each pixel and its spatial correlation information within the neighborhood, but also the dimension of neighborhood domain were encoded by genetic factors. The small range of the optimal threshold was gotten based on genetic algorithm, and in the small range, the global optimal threshold was found based on the second genetic algorithm computing. The improved algorithm was applied into cucumber computer vision system. The experiment results showed that, the consuming time of computing between-cluster variance of 2D maximum between-cluster variance algorithm based on two level genetic algorithm was 0.18% more than that of 2D maximum between-cluster variance algorithm, and was 46.87% more than that of Otsu algorithm. At the same time, the consuming time of computing between-cluster variance of 2D maximum between-cluster variance algorithm spent shorter time on running and had better segmentation effect than that of the traditional algorithms such as 2D maximum between-cluster variance algorithm and Otsu algorithm. The improved segmentation algorithm provides a new real-time image segmentation method for the object recognition field, so it has a certain promotion value.