Dragonfly-Net: Dragonfly Classification using Convolution Neural Network

Abstract

Scientific and engineering interests towards dragonflies has been a consistent source of ideas and solutions owing to the evolutionary success of the species. The importance of these ”toothed ones”, as the Greek translation of the family name ”Odonates” maps to, in terms of ecological diversity is invaluable, more pressingly with the context of only two of the six suborders of the order Odonata being non-extinct. With a wide spread existential timeline, identifying them is in itself is a critical task for taxonomists. This literature is oriented to provide a standard identification tool that aids researchers, amateur naturalists, and beginners in quick and easy identification of odonates, thus aiming to influence deeper exploration of the order. We propose a novel approach in terms of Dragonfly-Net, that has a widespread application possibility in the field of ecology and biology, starting with classification of a given image with dragonfly or dam-selfly into the pre-trained list of species belonging to the order, without any pre-processing. The proposed model performed with an accuracy of 76.99% on the training set, 67.59% on the validation set, and 61.35% on the hold-out set. The model predicts 94 different species of drag-onflies/damselflies. The effort is also protruded to derive and investigate the performance of the model with state-of-the-art evaluation techniques, scoped to explore the regions of activation contributing to its performance.

Keywords: Convolution neural networks, Deep learning, Dragonflies, Insect identification, Odonates.

View PDF