A Data Augmentation Approach Based on Generative Adversarial Networks for Date Fruit Classification

Highlights Classification of date fruit hardness levels from images. Combination of deep networks and an expert system to yield better accuracies. Abstract. Machine learning techniques have been used in various agricultural applications from farming to post-harvest operations. For some cases, a large amount of data is not available and improving on classification accuracies based on deep networks cannot be an option. Such is the case presented here for sorting date fruits based on their hardness into three classes (soft, semi-hard, hard). The original dataset in this work consists of 1800 monochrome images with 600 images per class obtained from different growing regions in Oman. This is a limited number of examples to consider deep networks. Thus, this work proposes data augmentation based on Generative Adversarial Networks (GAN) to synthetically augment the date fruit images. It incorporates a Densely Connected Convolutional Network (DenseNet) for date fruit classification. The goal is to generate enough images so that DenseNet can successfully classify them. The GAN images help during the training part when a large dataset is used for training. Accuracies of 100%, 100%, and 99% were achieved for the hard, soft, and semi-hard classes by using an expert system in combination with two DenseNet networks, one network trained for classifying hard and soft cases as well as a second network that classifies the three classes. Keywords: Data augmentation, Deep learning, Expert system, Fruit classification, GAN, Neural networks.