| Title |
Blind Image Quality Evaluation using Pseudo Reference Images |
| Authors |
용윤정(Yunjeong Yong) ; 오희석(Heeseok Oh) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.2.77 |
| Keywords |
Image quality assessment; Perceptual quality; Pseudo-reference image; Scale-invariant distortion; Quality score regressor |
| Abstract |
No-reference image quality assessment (NR-IQA) aims to objectively quantify the level of image quality degradation by reflecting the human visual system in the absence of information about the pristine image. Existing NR-IQA techniques have high sensitivity to specific distortion types, but have limitations in determining semantic quality information or global image quality degradation. In this paper, to resolve the limitations of existing NR-IQA approaches and improve predictive power, we propose a multi-scale pseudo image quality assessor (MPIQ). MPIQ is an NR-IQA model that follows the framework of full-reference IQA (FR-IQA), which is more proficient in extracting local distortion patterns and aggregating higher-level perceptual quality information. The proposed MPIQ employs a hybrid scheme that seeks to understand local distortion patterns through the convolutional neural networks and global level of image quality based on transformers, and consists of two modules: a pseudo-reference image reconstructor and a quasi FR-IQA regressor. Similar to the FR-IQA approach, the pseudo-reference image reconstructor utilizes an encoder-decoder structure to reconstruct the pseudo-reference image and learn image degradation information compared to a distorted one. Here, a multi-scale structure is reflected to extract scale-invariant distortion patterns. Quasi FR-IQA regressor predicts the image quality score by deriving the global distortion level through the difference between the features extracted from the distorted and the pseudo-reference images. MPIQ was supervised onto the mean opinion score obtained through subjective evaluation in an end-to-end manner, and experimental results showed a 20% performance improvement compared to the existing NR-IQA. |