Microscopic Fusarium detection and analyses with convolutional neural networks

Date of Award


Document Type


Degree Name

Master of Science in Computer Science


Information Systems & Computer Science

First Advisor

Estuar, Ma. Regina Justina E., Ph.D.


Recent advances in computer vision and artifcial intelligence, specifically deep convolutional neural networks, achieved state of the art results in multiple computer vision tasks providing possibilities in using these networks for the rapid detection of Fusarium oxysporum, the main cause of the Fusarium Wilt disease plaguing banana plantations. This study focused on the use and analysis of convolutional neural networks and deep learning by assessing multiple state ofthe art neural network architectures, such as Inception v3, MobileNet, ResNet,DenseNet. Extensive image augmentation techniques have been applied toinduce feature invariance in training the neural networks. Results have shown that Fusarium microconidia detection has been successful with fine tuning a MobileNet-based model trained from the ImageNet database with a combined F1score of 0.9702. Model analysis and verifcation with Locally Interpretable Modelagnostic Explanations, or LIME, and Gradient-weighted Class Activation Mapping,or Grad-CAM, had further narrowed down the specifc submodels trained from MobileNet to determine the behaviors of the models. These have shown that the models were able to successfully discriminate the microconidia of Fusarium oxysporum from other present artifacts, such as soil particles. Implications ofdata augmentation techniques have been discussed, specifcally on the efects of grayscale conversion.


The C7.L544 2018