Optimum CNN-based Plant Mutant Classification

Ng, Chee Ho (2018) Optimum CNN-based Plant Mutant Classification. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text NG CHEE HO - Full Thesis.pdf Restricted to Registered users only Download (2MB)

Abstract

This project aims to construct an optimum convolutional neural network (CNN) for plant mutant classification task. Tchebichef moments had been used to extract features from the plant images dataset. This project also explores the possibility to improve success classification rate by applying two image pre-processing methods: contrast adjustment and adaptive histogram equalization. Experiments had been designed and performed to find the optimum design of CNN for plant mutant classification. For first part, three parameters of a CNN had been experimented, which were optimum number of convolutional layer, optimum number of neuron for fully connected layer and optimum number of fully connected layer. After that, the optimum setting for CNN found in first part was used in second part. The optimum CNN model from first part was trained with three different dataset: raw RGB images, images pre-processed with contrast adjustment, and images pre-processed with adaptive histogram equalization. For third part, experiments were conducted to determine optimum design of CNN with Tchebichef moments as feature extractor. Image features had been extracted with Tchebichef moments order of 20 and order of 30. The obtained results showed that the optimum settings of CNN for plant mutant classification was using four convolutional layers, 4096 neurons for fully connected layer and two fully connected layers. The mean test classification accuracy for this optimum setting was 97.90%. Both image pre-processing techniques did not improve the classification accuracy. The mean test accuracy were 95.16% and 96.45% for pre-processing with contrast adjustment and pre-processing with adaptive histogram equalization, respectively. Plant mutant classification with Tchebichef moments showed encouraging results, with mean test accuracy of 94.52% and 95.80% for moment order number of 20 and 30 respectively. The contributions of this project is it presented a systematic method to determine the optimum setting of CNN for plant mutant classification task. Also, this project demonstrated the potential of using Tchebichef moments as a feature extractor for plant mutant classification task. The obtained results showed that it is capable of generating high classification accuracy while reducing the footprint of dataset size.