Setiawan, Wahyu Fajar (2026) Segmentasi Kerutan Wajah dengan Variasi Cahaya dan Pose Menggunakan Attention UNet-VGG dengan Augmentasi Geometrik dan Fotometrik. Masters thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 6025241020-Master_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
6025241020-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy |
Abstract
Segmentasi kerutan wajah penting dalam aplikasi dermatologi dan kosmetologi. Namun, implementasi dunia nyata menghadapi tantangan signifikan berupa variasi pose wajah dan kondisi pencahayaan tidak terkontrol, yang mengaburkan fitur tekstur halus dan menyebabkan penurunan akurasi segmentasi pada model konvensional. Penelitian ini bertujuan mengatasi permasalahan tersebut dengan mengembangkan metode segmentasi robust dan adaptif terhadap kedua faktor variasi secara simultan. Metode yang diusulkan mengintegrasikan arsitektur encoder-decoder dengan mekanisme attention. Tiga model dievaluasi: Attention UNet baseline, Attention UNet-VGG16, dan Attention UNet-VGG19. Multi-scale attention gates diterapkan agar model fokus adaptif pada area kerutan dengan variasi ketebalan, meminimalkan pengaruh background tidak relevan. Backbone VGG dipilih karena kemampuan superiornya mengekstraksi fitur tekstur hierarkis melalui transfer learning, krusial untuk mendeteksi pola kerutan halus. Model dilatih menggunakan combined loss function (Focal, Dice, IoU, dan Boundary Loss) dengan bobot optimal hasil Grid Search. Strategi augmentasi komprehensif menggabungkan transformasi geometrik (rotasi, shift, shear, zoom) dan multi-level enhancement fotometrik (CLAHE, koreksi gamma, bilateral filter, unsharp masking). Penelitian menggunakan dataset FFHQ-Wrinkle dengan 1000 citra wajah yang memiliki label manual, dibagi menjadi training (70%), validation (15%), dan testing (15%). Hasil eksperimen menunjukkan bahwa Attention UNet-VGG19 dengan kombinasi augmentasi mencapai performa terbaik dengan Dice Score 0.6533 dan IoU Score 0.4931, meningkat 4.26% dan 5.89% dibandingkan metode state-of-the-art terbaik (Striped WriNet) pada kondisi perbandingan yang adil. Model menunjukkan ketahanan tinggi dengan Robustness Score 0.6391, dimana degradasi performa pada kondisi kontras tinggi hanya 6.58%. Strategi kombinasi augmentasi meningkatkan Dice Score sebesar 13.24% dibandingkan baseline dan mereduksi overfitting gap dari 4.75% menjadi 0.19%. Temuan ini mengonfirmasi bahwa integrasi VGG encoder, attention mechanism, dan augmentasi komprehensif adalah solusi efektif untuk segmentasi kerutan yang robust.
===================================================================================================================================
Facial wrinkle segmentation is critical in dermatology and cosmetology applications. However, real-world deployment faces significant challenges from facial pose variations and uncontrolled lighting, which obscure fine texture features and cause accuracy drops in conventional models. This research aims to address these issues by developing a robust and adaptive segmentation method that handles both variation factors simultaneously.
The proposed method integrates encoder-decoder architecture with attention mechanism. Three models were evaluated: Attention UNet baseline, Attention UNet-VGG16, and Attention UNet-VGG19. Multi-scale attention gates enable adaptive focus on wrinkle areas with varying thicknesses, minimizing irrelevant background influence. VGG backbone was selected for its superior capability in extracting hierarchical texture features through transfer learning, crucial for detecting fine wrinkle patterns. Models were trained using combined loss function (Focal, Dice, IoU, Boundary Loss) with optimal weights via Grid Search. Comprehensive augmentation combines geometric transformations (rotation, shift, shear, zoom) and multi-level photometric enhancement (CLAHE, gamma correction, bilateral filter, unsharp masking).
The research utilizes the FFHQ-Wrinkle dataset comprising 1000 manually labeled facial images, split into training (70%), validation (15%), and testing (15%). Experimental results demonstrate that Attention UNet-VGG19 with combined augmentation achieves the best performance, with a Dice Score of 0.6533 and IoU Score of 0.4931, representing improvements of 4.26% and 5.89% compared to the best state-of-the-art method (Striped WriNet) under fair comparison conditions. The model exhibits high resilience with a Robustness Score of 0.6391, where performance degradation under high contrast conditions is only 6.58%. The combined augmentation strategy increased the Dice Score by 13.24% compared to baseline and reduced the overfitting gap from 4.75% to 0.19%. These findings confirm that the integration of VGG encoder, attention mechanism, and comprehensive augmentation constitutes an effective solution for robust wrinkle segmentation.
Actions (login required)
 |
View Item |