Synthesizing data for multiclass image classification

Image classification is a task used to identify what each individual image represents. Multiclass classification is a type of classification task which separates data into more than two classes. CNN, a class of deep neural network, is commonly used to analyse visual images. However, insufficient dat...

Full description

Saved in:
Bibliographic Details
Main Author: Lee, Tian Fa
Other Authors: Alex Chichung Kot
Format: Final Year Project
Language:English
Published: Nanyang Technological University 2020
Subjects:
Online Access:https://hdl.handle.net/10356/145163
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Image classification is a task used to identify what each individual image represents. Multiclass classification is a type of classification task which separates data into more than two classes. CNN, a class of deep neural network, is commonly used to analyse visual images. However, insufficient data will lead to lower accuracy in image classification. To elevate the performance of CNN, it is necessary to increase the amount of quality training data. Research have been done in DCGAN, but they did not evaluate whether different types of synthetic data are useful for data augmentation of images. In this project, a CNN model is proposed to evaluate the accuracy of different types of synthetic data and investigate their usefulness in multiclass classification. Due to limitation in computational power and memory, datasets with 10 classes such as MNIST and CIFAR-10 are used in the experiment. Different sets of tools such as DCGAN and pre-trained deep learning models and computer vision algorithms are used to generate DCGAN-original, cartoonized and sketched version of the dataset respectively. They are then added to the training dataset of the CNN with its test accuracy evaluated. After analysing the results, DCGAN-original data has the highest performance as its test accuracy is highest among the three. Furthermore, there is a higher potential to improve the accuracy of the DCGAN-original dataset by fine-tuning the DCGAN.