Herbal Plant Image Classification using Transfer Learning and Fine-Tuning Deep Learning Model

Abstract

Herbal plants are highly significant to the local community in Malaysia, as the country's fertile land is rich in diverse species that are widely used for various purposes, including traditional medicine, culinary, aromatherapy, and even in the cosmetic industries. This situation demands numerous applications and activities, including plant species identification, medicinal plant research, agriculture research, and environmental monitoring, which makes image classification of herbal plants a substantial task. However, this task is complicated by the complex nature of the plants, particularly includes plants variations in appearance, close similarities between species, and limited availability of labeled data and images.  Motivated to mitigate the issues, this paper investigated the use of transfer learning and fine- tuning the deep learning neural network to classify different herbal plant species. Transfer learning is an algorithm that learns to recognize image features in one domain and having the capability to generalize the learnt knowledge to a new domain with a smaller dataset. Additionally, fine-tuning can be used to further improve the performance of the model on the new task with less training time and fewer training data. The authors performed experiments on ResNet-50 which been previously trained with ImageNet dataset. The experiments were carried out on a subset of the MYLPherbs-1 dataset, which consisted of two local perennial herbs plant species. Different hyperparameters were used across the various experiment settings, and the authors observed the behavior and relationships of the distinct models, datasets, and hyperparameters toward the classification task's accuracy. The authors also employed two different transfer learning approaches: (i) using pre-trained models as feature extractors with different classifiers, (ii) fine-tuning the pre-trained model. Based on the results and discussion, fine-tuning the ResNet-50 model on the MYLPherb-1 dataset demonstrated the best overall performance.