In order to improve the citrus grading accuracy, fractal dimensions which characterize the color and shape features of citrus fruit were analyzed. Samples were from Citrus unshiu Marc.cv.unbergii Nakai. For each sample, images from peduncle, calyx and two opposite sides were collected. These four images were cut, removed backgrounds, and converted from RGB space to HSI one, then by the following methods, the color and shape features of citrus were extracted 1) HSI images were segmented according to hue value of 0°~20°, 20°~40°, 40°~60°, 60°~80°and 80°~100°. And each segment image was converted to binary image to retrieve box dimension, which character the color feature of fruit. 2) HSI images were converted to binary images, and then the imagines were edge detected and the box dimension of fruits profile of peduncle and one side, which character the shape features of fruits, were retrieved. Based on the box dimensions, a wavelet neural network was constructed to model the fruit color and shape grading system. Test results showed that for 120 sample fruits, the average correctness of color and shape grading was 95.83%, which mean box dimensions of equal segmentation of hue value 0°~100° revealed the color feature, and box dimensions of peduncle and side profile revealed fruit shape information. Color and shape grading accuracy meet the requirements for auto-grading of system real-time machine-vision.