Progressive Transfer Learning Approach for Identifying the Leaf Type by Optimizing Network Parameters

Abstract

There are many plant species with medicinal properties and hence, it becomes very crucial to recognize its relevance. With a range of plant species available for medicinal use, it becomes vital to classify them accurately for their efficient use in medicine. The precise and unerring classification of the plant species is beyond the reach of a common person since it requires thorough knowledge of the subject and manual recognition is tedious and inaccurate due to human error. The advocate solution engages the ReNet50 architecture for automated classification of medicinal leaf prior to the fine-tuning. The initial layers of the pre-trained ResNet50 model are freezed during the first phase of the training while the newly added layers are trained using a differential learning rate obtained by the one-cycle policy. The fine-tuned model from phase I is loaded and trained by unfreezing in the second phase. This process is repeated in such a way that the size of the image is made to increase progressively from 80, 128, 150 to 180 pixels in these two stages of learning. The proposed architecture is validated on four-leaf datasets, which also includes a self-created dataset of leaf images collected from the internet. The publicly available benchmark datasets of leaf images considered for carrying out experiments are Flavia leaf dataset, LeafSnap dataset (lab and field), and MalayaKew (MK) Leaf Dataset (D1 and D2). Several preprocessing techniques; such as identifying mislabeled and irrelevant images from the dataset, data preprocessing, and data balancing are applied before fine-tuning of the model in case of self-collected dataset from the internet. The top-1 accuracy of the Flavia dataset is 100%, whereas MK-D1 and MK-D2 datasets achieved top-1 accuracy of 99.05%, and 99.89% respectively. The Accuracy of the LeafSnap dataset is 97.95% and 99.43% for the field and lab images respectively.