About The Project • Built With • Download • Usage • License • Other Projects • Contact
Table of Contents
To further explore artificial intelligent systems within society, a cognitive neural network model that detects a disease in the cassava plant is taken and improved. Initial issues arise in the model’s runtime and GPU requirement allowing only a small number of epochs to happen, where theoretically, the more epoch the better the test accuracy. Outsourcing the GPU via a python web server on Google Col-lab and the runtime issue resolved, it was found that the system could have the potentially to improve the test accuracy up 67% given certain criteria met through training layer independently from each other. Through testing, it was proven that the more epoch ran, the better the test accuracy however the limit was beyond the GPU restrictions but could be seen in further development. The hindrances did not allow for the projects goals to be completed however sufficient improvement was made unto the existing model and is set for further development.
As a group we created various models and completed a variety of tests on these models with the main aim of improving accuracy for the testing set. We adjusted the epochs, we changed the shape of the model, the sizes of the images and augmented the images that were being read into the model. The aim was to find the best of all changes to eventually see how they worked together to end up with the best model we could for the Cassava data set, with a peak accuracy of 67% we feel that whilst it could be improved, for the size of the data set and the number of classifiers required of it that this result is very good. The images are complex and very large to start with so for a model with around 100 epochs to get a result like this was a good step forward and shows that perhaps with more changes to the data set, perhaps by flipping images along the vertical, we could have achieved an accuracy closer to 80% with more epochs to train the model. With the Cognitive Neural Network model, training the 2d convolutional layers separately from the dense layer allowed the system to train not just faster but it improved the test accuracy up to 63% which is the highest we’ve managed to get. The system however requires a lot of power to run the model and we confidently feel that the we have improved this model.
Python • pandas • tensorflow • matplotlib • opencv
The folder is ~1.0GB
Project Source: https://www.kaggle.com/competitions/cassava-leaf-disease-classification
Python Files: https://github.com/Mysftz/cassava-plant-disease-detection-classification
LaTeX Thesis: https://github.com/Mysftz/cassava-plant-disease-detection-classification-thesis
This project was submitted alongside a group dissertation for the degree of a Masters of Science in Computer Science (Artificial Intelligence) at the University of Kent in March 2022.
Distributed under the CC-BY-SA-4.0: Creative Commons Attribution Share Alike 4.0 International License. See LICENSE.txt for more information.
GitHub: @Mysftz · Portfolio: Website · LinkedIn: @lrgtomaszewski · Instagram: @Mysftz · Twitter: @MysftzUK