Plant Classification in the Wild: A Transfer Learning Approach

Abstract

Datasets specialized in wildlife usually contain imbalanced classes of natural wild images such as, for instance, plant images, which are acquired from the surrounding environment with natural scene background. Deep neural networks have proven their efficiency in classifying such datasets. However, such an approach requires a workaround to approximately balance the classes in order to prevent the occurrence of overfitting during the training phase of the neural network. Many approaches exist to overcome this problem includes over-sampling, undersampling, generating synthetic samples, data augmentation, etc. The iNaturalist species classification and detection dataset represents a good example of vastly imbalanced datasets. It contains 13 superclasses. This work focuses on the Plantae superclass and builds a Convolutional Neural Network to distinguish a subset of the subclasses of Plantae. Our model benefits from cutting-edge techniques such as transfer learning and data augmentation to obtain a reasonably high level of accuracy (78.76%).