Abstract: 【Objective】 The purpose of this project is to study a kind of dynamic grading method of walnut kernel based on machine vision technology that will be suitable for Xinjiang. 【Method】 Walnut kernel feature set was obtained from walnut image collected in real time and completed image preprocessing. Then mRMR feature selection algorithm was used to filter the original feature set and arrange the importance of the feature. Finally, support vector machine was used to analyze the importance of the feature. Three machine learning algorithms, decision tree and naive Bayes, were trained and tested, and the optimal classification method was obtained. Finally, the automatic tracking method and dynamic grading process of walnut kernel were designed, and the automatic classification system of walnut kernel was constructed. 【Result】 When using feature bin19, K1 and bin15 to train naive Bayesian classifier, the classification accuracy rate of walnut kernel classification was 97.33%. Under the dynamic condition, the walnut kernel automatic grading system was used to classify 150 walnut kernel,and the overall accuracy rate was 81.33%. 【Conclusion】 Based on the feature extraction and grading method of walnut kernel developed by machine vision, the method of walnut kernel dynamic grading can effectively complete the classification task of walnut color and integrity.