Publication

Abstract : Unmanned aerial vehicles (UAVs) are useful for acquiring images of epiphytes as they grow on other trees and in areas that are not easily accessible. Manually identifying epiphytes in these images is both time-consuming and prone to errors. Convolutional neural networks (CNNs) are the building blocks for almost all state-of-the-art image classification, detection, and segmentation tasks. The CNN algorithm generates good output results by using spatial information from the input images using different filters. U-Net is one of the widely used architectures for image segmentation with CNN building blocks. The transformers, also a state-of-the-art architecture proposed by Google, was designed for tasks with sequential data like sentence translation and text summarization in natural language processing (NLP). This architecture was designed using the concept of attention mechanism which helps in memorizing long sequences. In this study, we have used TransU-Net, an architecture that combines the merits of both transformer and CNN for the segmenting images of Werauhia kupperiana, an epiphyte, acquired with drones. The segmentation outputs generated from the trained models were evaluated with Dice score and Jaccard index.