Abstract: Aiming at the special planting requirements of garlic scale buds facing up and upright sowing, an automatic recognition algorithm of scale bud orientation with good practicability, high accuracy, and strong anti-interference was developed. This paper proposed an improved algorithm（CNN-SVM） based on convolutional neural network（CNN） and support vector machine（SVM） classification optimization to realize automatic identification and correction of garlic scale bud orientation. Additionally, SVM classification optimization scheme and random parameter selection was investigated, as well as a loss function detection method to solve the problems of small perception, poor classification effect, and over-fitting. The research results showed that the recognition accuracy of the CNN-SVM algorithm was 99.8%, and the recognition time of a single image was 0.024 s. Compared to the classic CNN and SVM algorithms, the proposed algorithm had a better effect on the recognition of small fields and strong interference, while simultaneously having advantages of high recognition accuracy, small calculation scale, and sensitivity to local features. This research not only provides algorithm reserves for the research and development of garlic automatic intelligent seeding equipment, but can also be promoted to other small object recognition.