서지주요정보
Diverse rare sample generation with pretrained GANs = 사전학습된 생성적 적대 신경망을 활용한 다양하고 희귀한 샘플 생성
서명 / 저자 Diverse rare sample generation with pretrained GANs = 사전학습된 생성적 적대 신경망을 활용한 다양하고 희귀한 샘플 생성 / Subeen Lee.
발행사항 [대전 : 한국과학기술원, 2025].
Online Access 원문보기 원문인쇄

소장정보

등록번호

8044929

소장위치/청구기호

학술문화관(도서관)2층 학위논문

MAI 25063

휴대폰 전송

도서상태

이용가능(대출불가)

사유안내

반납예정일

리뷰정보

초록정보

Deep generative models are proficient at generating realistic data but struggle to produce rare samples in low-density regions due to the scarcity of such samples in training datasets and the mode collapse problem. While recent methods aim to improve the fidelity of generated samples, they often reduce diversity and coverage by overlooking rare and novel instances. This study proposes a novel approach for generating diverse rare samples from high-resolution image datasets using pretrained GANs. Our method performs gradient-based optimization of latent vectors within a multi-objective framework and employs normalizing flows for density estimation in the feature space. This enables the generation of diverse rare images with controllable parameters for rarity, diversity, and similarity to a reference image. We demonstrate the effectiveness of our approach both qualitatively and quantitatively across various datasets and GAN architectures, without retraining or fine-tuning the pretrained models.

심층 생성 모델은 사실적인 데이터를 생성하는 데에는 능하지만, 훈련 데이터에 희귀 샘플이 부족하고 모델 학습 방식에 따른 모드 붕괴 문제가 존재하여, 밀도 낮은 영역에서는 표본을 효과적으로 생성하지 못한다. 최근 방법들은 생성된 각각의 표본의 품질을 향상하는 것을 목표로 하지만, 이는 희귀하고 새로운 표본을 무시함으로써 다양성과 커버리지를 저하하는 경우가 많다. 본 연구는 사전 학습된 생성적 적대 신경망(GAN)을 사용하여 고해상도 이미지 데이터 세트에서 다양한 희귀 표본을 생성하는 새로운 접근 방식을 제안한다. 제안하는 방법론은 다목적 최적화 구조 내에서 잠재 벡터의 그래디언트 기반 최적화를 사용하며, 특성 공간에서의 밀도 추정을 위해 정규화 흐름(Normalizing Flow) 모델을 활용한다. 표본의 희귀도, 다양성, 참조 이미지와의 유사성을 조절할 수 있는 파라미터를 이용하여 다양한 희귀 이미지를 생성할 수 있다. 우리는 사전 학습된 생성 모델을 재학습하거나 미세 조정하지 않고도 다양한 데이터 세트와 GAN에서 제안한 방법론이 효과적임을 질적 및 양적으로 입증하였다.

서지기타정보

서지기타정보
청구기호 {MAI 25063
형태사항 iv, 40 p. : 삽화 ; 30 cm
언어 영어
일반주기 저자명의 한글표기: 이수빈
지도교수의 영문표기: Jaesik Choi
지도교수의 한글표기: 최재식
Including appendix
학위논문 학위논문(석사) - 한국과학기술원 : 김재철AI대학원,
서지주기 References: p. 34-40
주제 Rare Sample
Deep Generative Model
GAN
Multi-Objective Optimization
Normalizing Flow
희귀 표본
심층 생성 모델
생성적 적대 신경망
다목적 최적화
정규화 흐름
QR CODE

책소개

전체보기

목차

전체보기

이 주제의 인기대출도서

Examples of diverse rare samples generated by our method.

Schematic diagram of our method.

Quantitative evaluation for Section 4.1.

Percentage of age, gender, race, and head pose attributes predicted by FaceXformer for Section 4.1.

Percentage of LFWA attributes predicted by FaceXformer for Section 4.1. Sorted in descending order of FFHQ(%). The entire table is in Table 6.5.

Examples of high- and low-density real and fake samples for Section 4.1.

Examples of rare samples generated by our method for Section 4.1.

Quantitative evaluation for Section 4.2.

Percentage of the cat-related breeds and dog-related groups in ImageNet classes predicted by Model Soups. Sorted in descending order of Real(%). The entire table is in Table 6.6.

Percentage of age, gender, race, head pose, Eyeglasses, and WearingHat attributes predicted by FaceXFormer. MetF. refers to the MetFaces dataset. The entire table is in Table 6.7.

Examples of high- and low-likelihood real samples from the AFHQ Cat, Dog, and MetFaces dataset.

Examples of rare samples generated by our method for Section 4.2.

Correlation plot for the k-NND / rarity score and negative log-likelihood estimated by the normalizing flow.

Example of the optimization path with a real k-NN manifold and a heatmap of likelihoods estimated by the normalizing flow model.

Rarity scores (RS) and LPIPS scores for the ablation study on the objective function.

Normalizing flow model architecture.

Rarity score and LPIPS score from our method with varying \(\lambda_{1}\) and \(\lambda_{2}\), experimenting on FFHQ-StyleGAN2. OOM refers to out-of-manifold sample percentage.

Rarity score and LPIPS scores from our method with varying \(k'\). LPIPS: Mean LPIPS score among the optimized samples. LPIPS*: Mean LPIPS score between reference and the optimized samples, experimenting on FFHQ-StyleGAN2. OOM refers to out-of-manifold sample percentage.

Examples of diverse rare samples generated by our method using FFHQ-StyleGAN2, varying the coefficients of the objective function, \(\lambda_1\)(similarity) and \(\lambda_2\)(diversity).

Rarity score and LPIPS score from our method with different distance metrics, experimenting on the FFHQ-StyleGAN2. OOM refers to out-of-manifold sample percentage.

Percentage of LFWA attributes predicted by FaceXFormer. The results are sorted in ascending order of FFHQ(%).

Examples of diverse rare samples generated by our method using FFHQ-StyleGAN2.

Additional rare samples generated by our method using FFHQ-StyleGAN2. In each row, the first and third columns serve as references for the second and fourth columns, respectively. The changed or generated attributes are listed below the figures. Rare attributes are highlighted in bold.

Comparative qualitative results: FFHQ-StyleGAN2 with a truncation value of \(\psi = 1.0\) (top-left), our method (top-right), and Polarity sampling with a truncation value of \(\rho = 1.0\) (bottom-left) and \(\rho = 5.0\) (bottom-right). Red boxes indicate out-of-manifold samples.

Percentage of cat breeds. The results are sorted in descending order of AFHQ Cat(%).

Percentage of LFWA attributes predicted by FaceXFormer. MetF. refers to the MetFaces dataset. The results are sorted in ascending order of MetF.(%).

Examples of diverse rare samples generated by our method using AFHQ and MetFaces with StyleGAN2-ADA.

Dog groups categorized from 120 dog classes in ImageNet dataset.

Additional rare samples generated by our method using AFHQ and MetFaces with StyleGAN2-ADA. In each row, the first and third columns serve as references for the second and fourth columns, respectively. The changed or generated attributes are listed below the figures. Rare attributes are highlighted in bold.

Examples of the optimization paths with a real k-NN manifold and a heatmap of likelihoods estimated by the normalizing flow. Top: The black line represents the optimization path, with each marker indicating a rarity score at every ten steps. The blue line represents the log-likelihood estimated by the NF model. Bottom: The balls indicate the nearby real k-NN manifold.

Quantitative evaluation of Polarity sampling for FFHQ and StyleGAN2 with varying \(\rho\). RS refers to the rarity score.

Quantitative evaluation of Polarity sampling with replacement for FFHQ and StyleGAN2.RS refers to the rarity score.

Qualitative results of Polarity sampling. Top: Sampling without replacement. Bottom: Sampling with replacement. Left: \(\rho = 1.0\). Right: \(\rho = 5.0\). Each set contains 20 randomly generated samples.

Qualitative results of Polarity sampling with \(\rho = 1.0\) (left) and \(\rho = 5.0\) (right), using online rejection sampling to prevent out-of-manifold samples. All samples are within the real k-NN manifold constructed from the FFHQ dataset.

High-likelihood references (top 100) and their optimized rare images generated by our method. Top: FFHQ-StyleGAN2. Bottom: MetFaces-StyleGAN2-ADA. Left: References with a truncation value of \(\psi = 1.0\). Right: Optimized images.

High-likelihood references (top 100) and their optimized rare images generated by our method. Top: AFHQ Cat-StyleGAN2-ADA. Bottom: AFHQ Dog-StyleGAN2-ADA. Left: References with a truncation value of \(\psi = 1.0\). Right: Optimized images.