Kuswardhani, Hajjar Ayu Cahyani (2025) Augmentasi Data Citra Mikroskopis Spermatozoa Manusia Menggunakan Wasserstein Generative Adversarial Network-Gradient Penalty (WGAN-GP). Undergraduate thesis, UPN Veteran Jawa Timur.
![[img]](https://repository.upnjatim.ac.id/style/images/fileicons/text.png) |
Text (Cover)
21083010044.-cover.pdf Download (1MB) |
|
Text (Bab 1)
21083010044.-bab 1.pdf Download (267kB) |
|
Text (Bab 2)
21083010044.-bab 2.pdf Restricted to Repository staff only until 2027. Download (700kB) |
|
Text (Bab 3)
21083010044.-bab 3.pdf Restricted to Repository staff only until 2027. Download (941kB) |
|
Text (Bab 4)
21083010044.-bab 4.pdf Restricted to Repository staff only until 2027. Download (667kB) |
|
Text (Bab 5)
21083010044.-bab 5.pdf Download (250kB) |
|
Text (Daftar Pustaka)
21083010044.-daftar pustaka.pdf Download (216kB) |
|
Text (Lampiran)
21083010044.-lampiran.pdf Restricted to Repository staff only until 2027. Download (250kB) |
Abstract
Manual sperm morphology analysis is hindered by its subjectivity and time-consuming nature, presenting a significant bottleneck in diagnosing male infertility. While deep learning models like YOLOv5 offer a promising solution for automation, their performance is highly contingent on the diversity of the training dataset. This study provides a systematic comparison of two data augmentation strategies: traditional geometric transformations and a generative approach using Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP). A YOLOv5 model was trained under three distinct scenarios: a baseline without augmentation, a model with traditional augmentation, and one supplemented with WGAN-GP-generated images. Our findings indicate that traditional augmentation proved to be the most effective strategy, yielding the highest mAP@0.5 of 0.595 with a strong balance between precision and recall. Although WGAN-GP augmentation improved performance over the baseline model (mAP@0.5: 0.539 vs. 0.494), the quality of its synthetic images, measured by a Fréchet Inception Distance (FID) score of 110, was insufficient to outperform traditional methods. We conclude that traditional geometric augmentation currently stands as the more robust and reliable method for optimizing object detection models for complex microscopic sperm imagery.
| Item Type: | Thesis (Undergraduate) | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Contributors: |
|
||||||||||||
| Subjects: | Q Science > QA Mathematics > QA76.6 Computer Programming Q Science > QA Mathematics > QA76.87 Neural computers |
||||||||||||
| Divisions: | Faculty of Computer Science > Departemen of Data Science | ||||||||||||
| Depositing User: | Hajjar Ayu Cahyani Kuswardhani | ||||||||||||
| Date Deposited: | 04 Dec 2025 04:35 | ||||||||||||
| Last Modified: | 04 Dec 2025 05:06 | ||||||||||||
| URI: | https://repository.upnjatim.ac.id/id/eprint/47165 |
Actions (login required)
 |
View Item |