Zero- and few-shot learning for diseases recognition of Citrus aurantium L. using conditional adversarial autoencoders

Abstract Plant diseases can cause significant production and economic losses, and also seriously restrict the sustainable development of agriculture. Traditional plant diseases recognition method is time-consuming and highly dependent on expert experience. Therefore, most of the existing works design models based on deep learning to automatic recognition. However, they are sample-intensive and hard for the diagnosis of some Citrus aurantium L. diseases with only a few or even zero labeled samples for training. In this paper, we propose a novel generative model for zero- and few-shot recognition of Citrus aurantium L. diseases using conditional adversarial autoencoders (CAAE). CAAE learns to synthesize visual features so that the zero- and few-shot recognition can be transformed to a conventional supervised classification problem. Specifically, CAAE consists of encoder, decoder, and discriminator. Different from conditional variational autoencoder (CVAE), we impose a discriminator to train the encoder by adversarially minimizing the loss between the prior distribution and the encoding distribution. Our model achieves a harmonic mean accuracy of 53.4% for zero-shot recognition of Citrus aurantium L. diseases, which is 50.4% higher than CVAE. Extensive experiments carried out on public zero-shot benchmark datasets and a further case study on our own collected dataset of Citrus aurantium L. diseases demonstrate that our model is suitable for the application of zero- and few-shot Citrus aurantium L. diseases diagnosis.