RM-IQA: A new no-reference image quality assessment framework based on range mapping method. (December 2021)

• Record Type: Journal Article
• Authors: Yuan, Tian; Li, Chen; Tian, Lihua; Li, Guo
• Is Part Of: Computers & electrical engineering. Volume 96:Part B(2021)
• Publication Date: 2021-12
• Subjects: Image quality assessment (IQA) -- Convolutional neural network -- Deep learning -- Range mapping; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2021.107508
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: A range-mapping framework is proposed to map an FR-IQA dataset into a new NR-IQA dataset. An end-to-end deep multi-task learning neural network is trained through a combination of datasets containing no-reference images and full-reference images. A pre-trained model without reference is utilized, which can greatly improve the accuracy and effect of the NR-IQA model. Abstract: Significant progress has been made in recent years in image quality assessment (IQA). In particular, the development of deep learning has provided no-reference (NR)-IQA with more impressive solutions. However, improving the generalization of NR-IQA models is still an urgent necessity. In this study, we propose a new framework that uses the range mapping method to map an existing full-reference (FR)-IQA dataset to an NR-IQA dataset, thereby further enhancing the accuracy and generalization of the NR-IQA model. First, an NR-IQA model is employed to score an FR-IQA dataset to obtain the corresponding mean opinion score (MOS) values. Then, the correlation coefficients between these MOS values and the original differential mean opinion score (DMOS) values marked by the FR-IQA dataset itself is calculated. Subsequently, the matching sequence pair is obtained according to these correlation coefficients. Then, a range mapping function is selected based on this sequence pair, and this function is used to map the entire FR-IQA dataset to the existing NR-IQA dataset, and a new NR-IQA dataset is generated. Finally, the new and the existing NR-IQA datasets are merged into a new dataset, which can train an end-to-end multi-task network to obtain the final model RM-IQA. This model exhibits better performance as it exploits more prior information. Based on the largest available NR-IQA dataset KonIQ-10k and FR-IQA dataset KADID-10K, the experimental results proved the effectiveness of the proposed framework. Graphical abstract: Image, graphical abstract