Effect of Annotation and Loss Function on Epiphyte Identification using Conditional Generative Adversarial Network

Abstract

The deep neural networks are capable of processing large amounts of data for image processing applications like classification, segmentation and identification etc. Conditional- Generative Adversarial Network (C-GAN) is an example of Deep Learning (DL)-based image processing algorithm that is used widely for identifying targets in images. In a Deep Learning based image processing application the quality of the input image plays an important role. The quality and variations within the dataset helps the neural network filters to derive the best features for the classification and identification of the target in the image. Many deep learning approaches will have a preprocessing pipeline which improves the quality of the input training data. The objective of this study was to evaluate the effect of preprocessing methods to improve C-GAN's ability to detect epiphytes in images acquired by Unmanned Aerial Vehicles (UAVs). These methods include 1) trimming the training images to include mostly the target plant, 2) generating annotation images with a thresholding technique, and 3) incorporating binary cross entropy loss function. Results obtained from this study shows that the percent occupancy of input images and annotation method plays an important role in identifying the target plant.