Abstract: Guangxi, a province in China, contributes more than 60% of China's sugarcane production per year. The sprouting rate of newly planted sugarcane is directly affected by viability of sugarcane buds. In sugarcane mechanized production, the detection of sugarcane bud viability for planting is still an issue and there is no a good classification method for differentiating viable sugarcane buds from the poor ones. To solve the problem, we proposed a classification method based on the minimum error rate of Bayes decision to detect the damaged buds as accurate as possible. Sugarcane cultivars of Taitang 22, Guitang 42 and Yuetang 60 were chosen for this study because they were representative cultivars of sugarcane in the area. We obtained sample images of sugarcane via image acquisition equipment, cut out the effective bud region images with computer vision technology, and extracted five features and eigenvalues including maximum gray scale. Then, a sample database was established, in which the image eigenvalues of the three sugarcane breeds was contained. Based on the data, prior probabilities of intact buds and damaged ones were calculated, from which eigenvalues' statistics of bud region images were obtained, and curve distribution diagrams of eigenvalues were described. By analyzing these statistics and diagrams, we found that all the distribution curves were connected end to end by horizontal lines and inclinable lines like zigzags. So these curves could not be fitted by normal distribution function, quadratic function, cubic function or exponential functions etc. Piecewise function could fit the distribution curve perfectly, but it would make the process of calculation and judgment more complicated when setting classification rules because of the excessive number of segments. Further observation showed that portions of small-proportion inclined lines and the saw tooth peak whose probability was zero could both be neglected. Then, the corresponding eigenvalues of the rest horizontal lines could be simplified to a uniform distribution. Calculating the even-distributed interval width and the entire interval width of the eigenvalues based on the database, the percentage between the two interval widths and the differences of percentage of the corresponding eigenvalues between the intact buds and damaged ones could be obtained. Using the percentage and the difference, even-distributed features could be confirmed. At the meantime, the conditional probability density was calculated by the probability density function of uniform distribution. Three features, including maximum gray scale, average gray scale and standard deviation of gray scale, which were the same even-distributed features of the three sugarcane cultivars, were chosen as the final classification features to simplify the classification rules. According to the full probability and Bayes formula, the prior probability and the conditional probability density of the final classification features were transformed to posterior probability, and then it could be classified by comparing the value of posterior probability. Matlab 2012b could be used to distinguish whether the bud was intact or not. The result of experiments showed that bud integrity classification accuracy of the three sugarcane cultivars were 92.09%, 93.49% and 93.02%, respectively. And the classification accuracy rate of damaged species had reached 98%, 97% and 96%, respectively. It proved that this classification method was feasible, which meant that it could detect damaged buds basically and provided a signal for eliminating the damaged buds automatically.