BananaGAN : Augmenting major banana disease detection using generated diseased pseudostem and rachis images

Early detection and timely management of crop diseases are essential for reducing yield loss. Traditional manual inspection is often time-consuming, laborious, and biased. Recently, automated imaging techniques have been successfully applied to the detection of crop diseases. Almost this type of research requires a huge amount of images with key typical symptoms from rare classes. The rare class images are the key to differentiated closely related diseased symptoms, but it is mostly internal and difficult to get them. Thus we exploited generative adversarial networks for generating rare classes such as banana pseudostem and rachis images creating new datasets with synthetic images and doing domain disease translation, converting an image with a certain disease into another image with another different disease. These synthetic images were tested in pre-trained disease detection models to see if they are good enough to balance the banana disease datasets and improve the object detection models’ overall accuracy and can be applied to other deep learning techniques such as classification and semantic segmentation. mAP score from the trained models with synthetic images was between 64% and 89% accuracy, which conclude that synthetic images are a useful tool as a data augmentation technique.

• Autores: Vergara, Javier Alejandro
• Tipo de recurso: Article of investigation
• Fecha de publicación: 2021
• Institución: Pontificia Universidad Javeriana Cali
• Repositorio: Vitela
• Idioma: eng
• OAI Identifier: oai:vitela.javerianacali.edu.co:11522/2021
• Acceso en línea: https://vitela.javerianacali.edu.co/handle/11522/2021
• Palabra clave: Artificial intelligence; Generative adversarial networks; Deep learning; Disease detection; Data augmentation; Pseudostem; rachis; Synthetic dat
• License: https://creativecommons.org/licenses/by-nc-nd/4.0/