Improvement of Classification Results of Convolutional Neural Networks Using Various Gan-Based Augmentation Techniques

Lenka Kališková, Peter Butka

Improvement of Classification Results of Convolutional Neural Networks Using Various Gan-Based Augmentation Techniques

Číslo: 4/2024
Periodikum: Acta Electrotechnica et Informatica
DOI: 10.2478/aei-2024-0013

Klíčová slova: augmentation, generative adversarial network, convolutional neural network, classification, deep learning

Pro získání musíte mít účet v Citace PRO.

Přečíst po přihlášení

Anotace: Generative Adversarial Network (GAN) is an exciting innovation in machine learning within the neural network field. These models are able to generate a realistic image, video or even voice output. One of the useful applications is its possibility to enrich data sets for better learning of neural network models. In the presented work, we focus on image augmentation with the use of several variations of GAN to improve the classification of convolutional neural network. Accordingly, to prove the advantage of GAN-based image augmentation in comparison with methods of classical augmentation, we used specifically three different degrees of image rotation and compared classification results of convolutional neural network that use images from these augmentation methods. Mentioned methods of image augmentation are applied to five datasets belonging to three different domains, specifically medical, astronomical and geological domain. The architecture and settings of the convolutional neural network are the same for all datasets. To evaluate classification results, we used confusion matrix, accuracy, precision, recall and F1-score.