Winter jujube has the characteristics of thin peel and crisp flesh, and winter jujube can only be picked by hand at present, so it is urgent to solve the problem of automatic and intelligent picking of winter jujube. Whereas, the recognition of winter jujube is the premise and foundation to solve this problem. In order to solve the problem of low recognition rate caused by the large number difference of samples with different levels of maturity, this paper proposes a recognition approach based on data-balanced Faster R-CNN for winter jujube. For the winter jujube with different levels of maturity in natural environment, this paper researches the Faster R-CNN recognition approach with data balance from different angles, and then the proposed method is compared with the recognition approach based on YOLOv3. The results show that: the proposed data-balanced Faster R-CNN method enhances the generalization effect of the model in the case of insufficient samples and unbalanced categories;the average recognition accuracy of the proposed approach is 98.50% which is higher than YOLOv3, and the total loss is less than 0.5. What’s more, the feature extraction of the foreground image is not obvious because the distance is far between the lens and the foreground image, which will reduce the recognition accuracy of the overall data set. This research has certain practical significance and application value for solving the recognition problem of automatic and intelligent picking winter jujube.