Citation

Liu, Hongyan and Jin, Shu and Gu, Ke and Shen, Xue and Wong, Lai Kuan and Lin, Weisi and Zhai, Guangtao and Zhang, Wenjun (2026) Fusing Quality-Aware Model and Large Language Model for Industrial Image Quality Assessment. IEEE Transactions on Circuits and Systems for Video Technology. p. 1. ISSN 1051-8215

Text 22.pdf - Published Version Restricted to Repository staff only Download (14MB)

Abstract

Industrial vision technology, as a core driving force in intelligent manufacturing, highly depends on the acquisition of high-quality image data, and thus it is strongly needed to devise an effective image quality assessment method for monitoring the quality of industrial image data.We can view industrial images as the compounds of natural-scene images that are commonly seen in our surrounding living environment and industrial elements that are hardly seen in our surrounding living environment but easily found from the internet. To this end, this work fuses the quality-aware model (QAM), which is developed to capture the difference of details and semantics based on the inspiration of quality assessment methods for natural-scene images, and the large language models (LLMs), which can acquire and analyze huge amounts of industrial images from the internet to assist in capturing the difference of the attributes of industrial elements, to propose a new method for assessing the quality of industrial images, referred to as QL-I2QA. First, this work devises a dualchannel QAM from two complementary viewpoints, including a “vision” channel that measures the difference of details between the reference and distorted images and a “brain” channel that measures the difference of semantics between the reference and distorted images. Second, this work constructs an “LLM” channel to measure the difference of the attributes of industrial elements between the reference and distorted images through elaborately designing the prompting method and response rules tailored to industrial attributes, thus to further improve the performance of the aforesaid QAM. Using two dedicated industrial image quality assessment databases, we tested and compared the proposed QL-I2QA method with more than two dozen classical and state-of-the-art competitors, and experimental results demonstrated that our QL-I2QA method achieved superior performance, markedly outperforming all the competitors.

Item Type:	Article
Uncontrolled Keywords:	Image quality assessment, industrial elements, quality-aware model (QAM), large language models (LLMs)
Subjects:	P Language and Literature > P Philology. Linguistics
Divisions:	Faculty of Computing and Informatics (FCI)
Depositing User:	Ms Suzilawati Abu Samah
Date Deposited:	05 Jun 2026 06:50
Last Modified:	05 Jun 2026 06:50
URII:	http://shdl.mmu.edu.my/id/eprint/16043

Downloads

Edit (login required)