No-Reference Image Quality Assessment Using Meta-Learning

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F62690094%3A18450%2F23%3A50020689" target="_blank" >RIV/62690094:18450/23:50020689 - isvavai.cz</a>

Výsledek na webu

<a href="http://dx.doi.org/10.1007/978-981-99-2680-0_13" target="_blank" >http://dx.doi.org/10.1007/978-981-99-2680-0_13</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1007/978-981-99-2680-0_13" target="_blank" >10.1007/978-981-99-2680-0_13</a>

Jazyk výsledku

angličtina

Název v původním jazyce

No-Reference Image Quality Assessment Using Meta-Learning

Popis výsledku v původním jazyce

Deep learning-based no-reference image quality assessment faces problems like dependency on a large amount of experimental data and the generalization ability of the learned model. A deep learning model trained on a specific dataset cannot obtain the desired results for testing on other datasets. Similarly, a deep learning model trained with small experiment data does not provide the best result. This paper addresses these problems of the deep learning model using the meta-learning approach in the field of no-reference Image Quality Assessment. The no-reference image quality assessment is a small sample problem, where the amount of experimental data is very less. Although data augmentation techniques have been used to increase the amount of experimental data, they do not increase the variation of the data. Therefore, traditional deep learning-based techniques are unsuitable for no-reference image quality assessment. Another problem is the lack of generalization ability. A deep learning model trained with image quality datasets containing images distorted synthetically cannot efficiently assess the quality of images distorted naturally. This work proposes a meta-learning model that can be trained with more than one image quality dataset, where one image quality assessment dataset contains synthetic images and the other contains real images. Finally, another image quality assessment dataset has tested the trained model. The test result is better than the state-of-the-art methods, and the results establish the fact that the meta-learning model proposed in this paper tries to resolve the problems of the deep learning model. © 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.

Název v anglickém jazyce

No-Reference Image Quality Assessment Using Meta-Learning

Popis výsledku anglicky

Deep learning-based no-reference image quality assessment faces problems like dependency on a large amount of experimental data and the generalization ability of the learned model. A deep learning model trained on a specific dataset cannot obtain the desired results for testing on other datasets. Similarly, a deep learning model trained with small experiment data does not provide the best result. This paper addresses these problems of the deep learning model using the meta-learning approach in the field of no-reference Image Quality Assessment. The no-reference image quality assessment is a small sample problem, where the amount of experimental data is very less. Although data augmentation techniques have been used to increase the amount of experimental data, they do not increase the variation of the data. Therefore, traditional deep learning-based techniques are unsuitable for no-reference image quality assessment. Another problem is the lack of generalization ability. A deep learning model trained with image quality datasets containing images distorted synthetically cannot efficiently assess the quality of images distorted naturally. This work proposes a meta-learning model that can be trained with more than one image quality dataset, where one image quality assessment dataset contains synthetic images and the other contains real images. Finally, another image quality assessment dataset has tested the trained model. The test result is better than the state-of-the-art methods, and the results establish the fact that the meta-learning model proposed in this paper tries to resolve the problems of the deep learning model. © 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

—

Návaznosti

S - Specificky vyzkum na vysokych skolach

Rok uplatnění

2023

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

Lecture Notes in Networks and Systems

ISBN

978-981-9926-79-4

ISSN

2367-3370

e-ISSN

—

Počet stran výsledku

8

Strana od-do

137-144

Název nakladatele

Springer Science and Business Media Deutschland GmbH

Místo vydání

Singapur

Místo konání akce

Ropar

Datum konání akce

19. 12. 2022

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—