Unsupervised Image Translation using Adversarial Networks for Improved Plant Disease Recognition

Acquisition of data in task-specific applications of machine learning like\nplant disease recognition is a costly endeavor owing to the requirements of\nprofessional human diligence and time constraints. In this paper, we present a\nsimple pipeline that uses GANs in an unsupervised image translation environment\nto improve learning with respect to the data distribution in a plant disease\ndataset, reducing the partiality introduced by acute class imbalance and hence\nshifting the classification decision boundary towards better performance. The\nempirical analysis of our method is demonstrated on a limited dataset of 2789\ntomato plant disease images, highly corrupted with an imbalance in the 9\ndisease categories. First, we extend the state of the art for the GAN-based\nimage-to-image translation method by enhancing the perceptual quality of the\ngenerated images and preserving the semantics. We introduce AR-GAN, where in\naddition to the adversarial loss, our synthetic image generator optimizes on\nActivation Reconstruction loss (ARL) function that optimizes feature\nactivations against the natural image. We present visually more compelling\nsynthetic images in comparison to most prominent existing models and evaluate\nthe performance of our GAN framework in terms of various datasets and metrics.\nSecond, we evaluate the performance of a baseline convolutional neural network\nclassifier for improved recognition using the resulting synthetic samples to\naugment our training set and compare it with the classical data augmentation\nscheme. We observe a significant improvement in classification accuracy (+5.2%)\nusing generated synthetic samples as compared to (+0.8%) increase using classic\naugmentation in an equal class distribution environment.\n