Trans-SURNet: A linear transformer approach to model picture-wise JND distribution for SUR curve prediction

TY - JOUR

T1 - Trans-SURNet

T2 - A linear transformer approach to model picture-wise JND distribution for SUR curve prediction

AU - Zhang, Anni

AU - Pei, Laifan

AU - Fan, Chunling

AU - Lin, Hanhe

N1 - Publisher Copyright: © 2026 Elsevier Inc.

PY - 2026/8

Y1 - 2026/8

N2 - The picture-wise just noticeable difference (PJND) reveals the smallest distortion level that is perceived by a user when comparing a reference pristine image to its distorted versions. The satisfied user ratio (SUR) curve for a lossy image compression scheme, e.g., JPEG, corresponds to the complementary cumulative distribution function of the PJND ratings. Although accurately and promptly predicting SUR curve is beneficial to many real-world applications, the existing methods are inefficient. In this paper, we propose a novel model named Trans-SURNet. Different from existing methods that predict the SUR value for a given reference image and its distorted version separately, our method predicts the probability density function (PDF) of PJND ratings for all distorted versions of a given reference image in an end-to-end manner, from which the SUR curve can be easily derived, thereby improving efficiency. Moreover, our method improves the performance by learning quality-aware features and adapting the proposed linear Transformer variant entitled Koalaformer to model how the change of distortion impacts the PDF of PJND ratings. Extensive experiments on two PJND benchmark datasets demonstrate the efficiency and effectiveness of our proposed method. The source code of the Trans-SURNet model is available at https://github.com/wwwzan/Trans-SURNet.

AB - The picture-wise just noticeable difference (PJND) reveals the smallest distortion level that is perceived by a user when comparing a reference pristine image to its distorted versions. The satisfied user ratio (SUR) curve for a lossy image compression scheme, e.g., JPEG, corresponds to the complementary cumulative distribution function of the PJND ratings. Although accurately and promptly predicting SUR curve is beneficial to many real-world applications, the existing methods are inefficient. In this paper, we propose a novel model named Trans-SURNet. Different from existing methods that predict the SUR value for a given reference image and its distorted version separately, our method predicts the probability density function (PDF) of PJND ratings for all distorted versions of a given reference image in an end-to-end manner, from which the SUR curve can be easily derived, thereby improving efficiency. Moreover, our method improves the performance by learning quality-aware features and adapting the proposed linear Transformer variant entitled Koalaformer to model how the change of distortion impacts the PDF of PJND ratings. Extensive experiments on two PJND benchmark datasets demonstrate the efficiency and effectiveness of our proposed method. The source code of the Trans-SURNet model is available at https://github.com/wwwzan/Trans-SURNet.

KW - Image compression

KW - Linear transformer

KW - Picture-wise just noticeable difference

KW - Satisfied user ratio

KW - Transmission reduction

UR - https://www.scopus.com/pages/publications/105043226190

U2 - 10.1016/j.jvcir.2026.104867

DO - 10.1016/j.jvcir.2026.104867

M3 - Article

AN - SCOPUS:105043226190

SN - 1047-3203

VL - 119

JO - Journal of Visual Communication and Image Representation

JF - Journal of Visual Communication and Image Representation

M1 - 104867

ER -