System development of image data augmentation using generative adversarial network (GAN)

Lim, Hou Hua (2022) System development of image data augmentation using generative adversarial network (GAN). Project Report. Universiti Teknikal Malaysia Melaka, Melaka, Malaysia. (Submitted)

	Text (24 Pages) System development of image data augmentation using generative adversarial network (GAN).pdf - Submitted Version Download (1MB)
	Text (Full text) System development of image data augmentation using generative adversarial network (GAN).pdf - Submitted Version Restricted to Repository staff only Download (2MB)

Abstract

A project is proposed to investigate the use of Generative Adversarial Networks (GAN) to augment image data to overcome class imbalance problems in silicon wafer microcrack defect classification. This class imbalance problem is due to the small number of defective products compared to non-defective products in a production environment. A Mask GAN is used to generate images of defect mask label when bounding box information is supplied. The artificial defect masks are used by a Defect Image GAN to “paint” defects onto non-defect images. The system's feasibility is analyzed by evaluating the accuracy of the classifier trained on the augmented data, along with the number of misclassified images of defective products. Two identical classifier networks are trained on traditionally augmented data and GAN-augmented data and be compared. The system is shown with McNemar’s test with a confidence of 95%, to produce a significant improvement over conventional methods.

Actions (login required)

View Item