Microscopic Fusarium Detection and Verification with Convolutional Neural Networks

Document Type

Conference Proceeding

Publication Date



Advances in computer vision, specifically on deep convolutional neural networks, have achieved state of the art results in multiple computer vision tasks. These networks have enabled the rapid de-tection of Fusarium oxysporum, the main cause of the Fusarium Wilt disease plaguing banana Cavendish plantations. The paper focuses on the use of convolutional neural networks and deep learn-ing by training and fine-tuning a MobileNet-based deep learning model for Fusarium detection in microscopy images. Multiple im-age augmentation techniques have been used to induce feature invariance in the model to control for the unique characteristics of Fusarium. Analysis of the behavior of the trained model using Locally Interpretable Model-agnostic Explanations, or LIME, has been performed to verify correct behavior. Results with MobileNet and LIME have shown that two out of four models have been able to specifically discriminate Fusarium from other present artifacts, such as soil particles.