Omnidirectional Image Quality Captioning: A Large-Scale Database and a New Model.

Area of Science:

• Computer Vision
• Image Processing
• Human-Computer Interaction

Background:

• Omnidirectional images are increasingly used, necessitating robust quality assessment methods.
• Current omnidirectional image quality assessment (OIQA) models struggle with heterogeneous distortions.
• Existing methods are often evaluated on limited datasets with homogeneous distortions.

Purpose of the Study:

• To establish a large-scale database (OIQ-10K) for evaluating omnidirectional image quality assessment (OIQA) under diverse distortion types.
• To develop a novel, adaptive OIQA model capable of handling both homogeneous and heterogeneous distortions in omnidirectional images.
• To investigate the relationship between subjective quality, objective distortion metrics, and human visual attention in omnidirectional images.

Main Methods:

• Creation of the OIQ-10K database, comprising 10,000 omnidirectional images with varied distortions.
• Conducting a comprehensive psychophysical study to gather subjective quality scores, distortion spatial distributions, and eye-tracking data.
• Development of IQCaption360, a multitask-derived adaptive feature-tailoring model for generating image quality captions.

Main Results:

• The OIQ-10K database provides a benchmark for challenging OIQA tasks.
• The proposed IQCaption360 model demonstrates superior performance compared to state-of-the-art methods on the OIQ-10K database.
• The model effectively generates textual quality captions, indicating its practical applicability.

Conclusions:

• IQCaption360 offers a significant advancement in omnidirectional image quality assessment, particularly for complex distortion scenarios.
• The OIQ-10K database and IQCaption360 model are valuable resources for future research in OIQA.
• This work highlights the importance of considering heterogeneous distortions and human perception in image quality evaluation.