Identifying strawberry appearance quality by vision transformers and support vector machine

Abstract The strawberry is a highly nutritional and beneficial cash crop, and its appearance quality sorting is a crucial step in the production process. Manual identifying and sorting, however, are subjective and more time consuming. In the field of image classification, vision transformers (ViT) have shown better performance. In this article, an extensive evaluation of the ViT models for the identification of strawberry appearance quality is presented. Moreover, to balance the accuracy and computation time of the ViT model, we propose a ViT‐based method that uses fine‐tuned ViT‐B/32 to extract the class token and imports it into the support vector machine (SVM) to identify strawberry. Besides, the class token is imported into the SVM with different kernel functions to evaluate their performance. The experimental results show that the highest recognition accuracy of original ViT models is ViT‐L/16, which can reach 97.38%. The application of the linear kernel function is more suitable in this work. The accuracy of original ViT‐B/32 is 93.7% and the proposed method improves the accuracy by 4.4% up to 98.1%. Furthermore, the required computation time of the proposed method is only 62.13 s, which is faster than other models. Therefore, the proposed method demonstrates enhanced robustness and universality, especially for abnormal and ripe strawberries. Practical Applications The identification of strawberry appearance quality is a labor‐intensive task. Traditional methods mainly rely on manpower, which has high cost and low efficiency. Therefore, this work uses the fine‐tuned ViT‐B/32 to extract features and import them into SVM to identify the appearance quality of strawberries, which obtains the identification results more accurate and efficient. This study could facilitate the development of smart picking equipment for strawberries.