Deep-learning based detector for identification of woodiness disease in passion fruits
Abstract
In this project, an AI-based algorithm based on a combination of Faster Regional Convolutional Neural Network (F-RCNN) and Residual Network combination framework was developed. The algorithm detects woodiness, a passion fruit disease in presence of the other classes of diseases. Several farms in Hoima and Luwero were visited and manually collected over 10,000 images for the dataset using a 24MP NIKON D5300 with an APS-C sensor camera. With the framework, images were preprocessed that is resized, labeled, classifi ed and then divided into \Test" and \Train" folders. \Train" folder images were used learn our labeled
features over 200,000 iterations a single graphical processor (GPU), the NVIDIA-1080TI.
\Test" images were input to the framework for classifi cation and detection and testing.
The algorithm performs well despite few training iterations, illumination variations, pose, size and orientation. Woodiness disease was successfully detected at an accuracy of 97 % under this limited dataset.