Swin-MLP: a strawberry appearance quality identification method by Swin Transformer and multi-layer perceptron

Accurate identifying of strawberry appearance quality is an important step for robot picking in the orchard. The convolutional neural network (CNN) has greatly helped the computer vision tasks such as the identification of fruits. However, better performance of CNN requires more time and computation for training. In order to overcome these shortcomings, a method, named “Swin-MLP”, based on Swin Transformer and multi-layer perceptron (MLP) to identify the strawberry appearance quality is proposed. The proposed method utilizes the Swin Transformer to extract strawberry image features and then import the features into MLP for identifying strawberry. In addition, the performance of combinations of Swin Transformer plus diffident classifiers is evaluated. Furthermore, the proposed Swin-MLP method is compared with original Swin-T and traditional CNN models. The accuracy of the proposed method reaches 98.45%, which is 2.61% higher than original Swin-T model. The required training time of the Swin-MLP only is 16.79 s that is extremely faster than other models. The experiment results show that the Swin-MLP has a good effect on identifying strawberry appearance quality. The success of the proposed method provides a new solution for strawberry quality identification.