Aiming at counting the grain number per spike in wheat breeding, a novel method based on fractal image segmentation is proposed in this paper. Firstly, a square image window is selected. Then, the fractal dimension of the window is set as eigenvalue of its center pixel. In order to extract the wheat spike, a proper threshold is selected to segment the image according to the pixel eigenvalue. To make the wheat spike perpendicular to the row direction of image, the angle between the wheat spike and the vertical direction of the segmented image is computed through analyzing row data of the image. The image is rotated according to the angle, and an angle-adjusted image is obtained. At last, the grain number can be calculated by column data of the angle-adjusted image. Experiments show this method achieves a higher accuracy and better efficiency than the traditional counting method.