A Lightweight Flower Classification Model Based on Improved Knowledge Distillation

Fast and accurate flower identification is significant in practice for botany and agriculture. Traditional flower classification models usually increase the parameters and computation burden to improve the accuracy, which is slow and impractical to be applied in low computing power devices such as mobile. To solve this problem, a flower classification model is constructed by lightweight convolutional neural network in this paper. To further improve the accuracy of the model, an improved method of knowledge distillation based on temperature decay (EDKD) is proposed by the authors, and apply it to the training of lightweight flower classification model. Experimental results on the Oxford102 flower dataset show that the model achieves an accuracy of 98.76% without increasing parameters or enlarging size, which outperforms the conventional methods and is more suitable for the mobile end.