PhD defense of Yujin CHO

Yujin CHO

On Blind Image Quality Assessment: From Camera Pipeline to Perceptual Evaluation

Abstract

Smartphone image quality depends on both the Image Signal Processing (ISP) pipeline that transforms RAW sensor data into display-ready images and the perceptual quality of the resulting rendering. However, developing and evaluating these systems remains challenging due to the limited availability of aligned RAW–RGB pairs, reliable reference images, and expert perceptual annotations. This thesis was carried out as part of a CIFRE collaboration between Centre Borelli at École Normale Supérieure Paris-Saclay and DXOMARK. It addresses these challenges in two parts. The first part focuses on learning and evaluating single-frame ISP pipelines with limited reference data. It first introduces an unpaired RAW-to-RGB learning method that constructs pseudo-pairs from semantically related images, enabling a compact model to reproduce a target camera's rendering without aligned training data. It then proposes a reference-free evaluation framework that estimates conventional full-reference quality criteria from processed images and capture information, while efficiently adapting to unseen ISP configurations. The second part investigates perceptual portrait quality assessment. Since expert annotations are costly and difficult to scale, the first study efficiently adapts a pretrained image quality model to predict expert judgments of texture preservation and noise quality using limited portrait-specific annotations. The second study examines facial contrast, a perceptually important but insufficiently defined attribute, by decomposing it into interpretable components and developing dedicated metrics validated against human judgments. Overall, this thesis provides practical and interpretable approaches to ISP learning, objective evaluation, and perceptual portrait assessment, while reducing reliance on costly reference data and extensive subjective evaluations.

PhD surpervisors

• Gabriele Facciolo, Professeur, Centre Borelli, ENS Paris-Saclay
• Jean-Michel Morel, Professeur, Lingnan University

Jury members

•  Yann Gousseau, Professeur, Télécom Paris (Rapporteur)
• Julien Mairal, Directeur de recherche, Inria Grenoble, Université Grenoble Alpes (Rapporteur)
• Jean Ponce, Professeur, École normale supérieure et New York University (Examinateur)
• Lara Raad Cisa, Professeure assistante, IIE, FIng, Universidad de la República (Examinatrice)

Invited Jury Members

• Benoît Pochon, Ingénieur, DXOMA
• Sira Ferradans, Chercheuse, DXOMARK