Multi-inference strategy for self-supervised denoising problem = 자기지도 디노이즈 문제를 위한 다중 추론 전략

국문 초록 (Abstract)

자체지도이미지노이즈제거는깨끗한실제값의지도없이노이즈이미지 세트를 사용하여 깨끗한 이미지 재구성을 목표로 하는 ill pose 문제이다. 실제 값에 과대한 의존은 과적합 및 분산 감소와 같은 문제를 초래할 수도 있어서, 현대 접근 방식은 특정 데이터 확대 또는 정규화 기술을 사용하여 이러한 문제들을 우회한다. 그러나 이런 접근 방식은 여전히 다양한 영역에서 일반화된 성능을 달성하지 못하고 있다. 게다가, 이런 전략들은 노이즈제거와 같은 분야에서 추가적인 하이퍼 파라미터에 의존한다. 최적의 하이퍼 파라미터 추정은 어려우며 파라미터를 잘못 조정하면 성능 저하를 불러일으킬 수 있는 oversmoothing 및 일관성 없는 이미지 구조가 발생할 수 있다. 본 논문에서는 앞서 서술한 문제점을 완화하기 위한 파라미터에 의존하지 않는 self-regularization 기법을 제안한다. 다중 추론 전략에서는 다양한 이미지 예측 결과를 정규화로
활용하여 콤팩트한 손실 함수를 정의한다. 더불어, 제안된 자체 정규화 방법은 모든 데이터 확대 기술과 모델 의존성에서 의존하지 않고 간단한 훈련 전략을 사용한다. 다양한 실험을 통해 제안하는 손실함수는 합성노이즈 영역 및 실제 노이즈 영역에서도 기존 기법들과 성능 우위 격차를 보여주고 있다.

다국어 초록 (Multilingual Abstract)

Self-supervised image denoising is a challenging problem that aims at signal reconstruction on a sparse set of noise measurements without any supervision of clean ground truths. Conventional supervised methods consider the noise recovery process as an ill-posed optimization problem with the availability of ground truth which is challenging in numerous domains. Self-supervised techniques alleviate the ground truth-unavailability issue by incorporating several complicated objective functions for proper noise removal and reconstruction.
However, the diverse noise distribution of images is crucial for noise recovery. Moreover, to form a complex loss function, the methods need to rely on additional hyperparameters. However, optimal hyperparameter estimation is complicated, and any mistuning of the parameters results in over-smoothing and inconsistent
structure recovery that is responsible for performance degradation. This paper proposes a self-regularization technique without using any hyperparameter to alleviate the aforementioned issues. Our multiple predictions acquired from a multi-inference self-supervised strategy are exploited as the regularization parameters and produce a compact loss function. Moreover, the proposed selfregularized method achieves satisfactory performance using multiple models and follows a simple training strategy without any complexity. Our experimental results represent that our compact loss function can achieve satisfactory performances in comparison to other existing methods for both synthetic and real noise domains. We also implement our algorithm on practical applications to represent how such low-level vision task is effective in high-level vision
applications.We represent a comparison scenario with weakly and un-supervised denoising methods to highlight our improved performance in the above applications.

목차 (Table of Contents)

• Ⅰ. Introduction 1
• A. Conventional supervised learning in image denoising 2
• B. Self-supervised learning in image denoising 4
• C. Motivations 7
• D. Contributions 9
• E. Thesis Layout 10
• Ⅱ. Background 11
• A. From supervised to self-supervised 11
• 1. Blind-spot based methods. 12
• 2. Unblind methods. 14
• B. Self-regularization effect on loss-function estimation. 15
• Ⅲ. Related Studies 17
• A. Non-learning based image denoisers 17
• B. Supervised learning with paired noisy/clean version. 17
• C. Denoisers trained with pairs. 18
• D. Unsupervised denoisers. 18
• E. Self-supervised denoisers. 18
• Ⅳ. Proposed Framework. 21
• A. Intuition. 21
• B. Mathematical justification. 22
• C. Multi-inference strategy. 25
• D. Loss function. 27
• Ⅴ. Experiments. 31
• A. Training details. 31
• B. Synthetic noise removal experiments. 32
• C. Results of Synthetic Experiments. 33
• 1. Gaussian Noise removal result. 37
• 2. Poisson noise removal. 40
• D. Real-noise removal experiments. 42
• E. Experiments on CC and PolyU. 45
• Ⅵ. Applications. 50
• A. Multiface detection. 50
• B. Object detection. 52
• Ⅶ. Conclusion. 55