THE EFFECT OF DCGAN-BASED DATA AUGMENTATION ON THE PERFORMANCE OF EFFICIENTNET-B0 AND MOBILENETV2 IN ANIMAL IMAGE CLASSIFICATION

Pasaribu, Fieter Brain (2026) THE EFFECT OF DCGAN-BASED DATA AUGMENTATION ON THE PERFORMANCE OF EFFICIENTNET-B0 AND MOBILENETV2 IN ANIMAL IMAGE CLASSIFICATION. Undergraduate thesis, Universitas Pendidikan Ganesha.

[img] Text (ABSTRAK)
2215101018-ABSTRAK.pdf
Download (237kB)
[img] Text (BAB 1 PENDAHULUAN)
2215101018-BAB 1 PENDAHULUAN.pdf
Download (216kB)
[img] Text (BAB 2 KAJIAN TEORI)
2215101018-BAB 2 KAJIAN TEORI.pdf
Download (539kB)
[img] Text (BAB 3 METODELOGI PENELITIAN)
2215101018-BAB 3 METODELOGI PENELITIAN.pdf
Restricted to Repository staff only
Download (611kB) | Request a copy
[img] Text (BAB 4 HASIL DAN PEMBAHASAN)
2215101018-BAB 4 HASIL DAN PEMBAHASAN.pdf
Restricted to Repository staff only
Download (2MB) | Request a copy
[img] Text (BAB 5 PENUTUP)
2215101018-BAB 5 PENUTUP.pdf
Restricted to Repository staff only
Download (185kB) | Request a copy
[img] Text (DAFTAR PUSTAKA)
2215101018-DAFTAR PUSTAKA.pdf
Download (227kB)
[img] Text (LAMPIRAN)
2215101018-LAMPIRAN.pdf
Download (404kB)
[img] Text (COVER)
2215101018-COVER.pdf
Download (568kB)

Abstract

Deep learning classification models, such as Convolutional Neural Networks, still face challenges due to their reliance on large datasets. When the available data is limited, the models are prone to overfitting, meaning they achieve strong performance on training data but fail to generalize well to unseen test data. To overcome this issue, this study investigates the use of synthetic data augmentation through Deep Convolutional Generative Adversarial Networks (DCGAN). This research evaluates the effect of synthetic data on the performance of EfficientNetB0 and MobileNetV2 in classifying four animal categories from the STL10 dataset. Synthetic data were generated using DCGAN at varying ratios, from 0.5 to 20 times the original data, and combined with real training data. Model performance was assessed using accuracy, precision, recall, F1 score, and confusion matrix visualization. The results show that EfficientNetB0 consistently outperformed MobileNetV2, achieving the highest test accuracy of 77 %. While DCGAN helped expand the dataset, the synthetic data still lacked realism and diversity, as indicated by relatively high FID and moderate IS scores. In some cases, adding synthetic data reduced accuracy, especially when poor-quality images dominated. There is no universally optimal ratio of synthetic to real data, ideal results depend on the dataset, model, and image quality. In conclusion, DCGAN can be a useful method for addressing data scarcity in image classification tasks, but its application must be carefully validated to avoid degrading model performance. Future research is recommended to enrich and diversify the dataset to enhance model generalization and image quality. In addition, more advanced GAN architectures such as WGAN-GP, StyleGAN, ProGAN, or BigGAN should be explored.

Item Type: Thesis (Undergraduate)
Uncontrolled Keywords: DCGAN. klasifikasi gambar, data sintetis, EfficientNetB0, MobileNetV2, STL10
Subjects: T Technology > T Technology (General)
Divisions: Fakultas Teknik dan Kejuruan > Jurusan Teknik Informatika > Program Studi Ilmu Komputer (S1)
Depositing User: Fieter Brain Pasaribu
Date Deposited: 02 Jun 2026 00:24
Last Modified: 02 Jun 2026 00:24
URI: http://repo.undiksha.ac.id/id/eprint/29294

Actions (login required)

View Item View Item