Automated Flower Classification using Transfer Learning and Meta-Learners in Deep Learning Framework

Abstract

The classification of flowers is a challenging task due to the wide variety of flowers along with inter-and intra-variations amongst the flower categories. Furthermore, the information such as the grass and leaves does not help in providing context to the recognition system. Researchers have extensively used deep learning frameworks for improving the classification accuracy but there is still room for improvement in terms of the recognition performance. In this paper, we use the transfer learning aspect to fine-tune the existing pre-trained networks which provide us an edge for the improved classification accuracy. We then apply various decision-level fusion strategies to combine the class probabilities from the individual pre-trained networks for further boost in recognition performance. Our method has been validated on two well-known flower datasets. The experimental results show that the proposed method achieves the best performance i.e. 99.80 % and 98.70 % on Oxford-17 and Oxford-102 datasets, respectively, which is better than the state-of-the-art methods.