LFACon: Introducing Anglewise Attention to No-Reference Quality Assessment in Light Field Space

Area of Science:

• Computer Vision
• Image Processing
• Virtual Reality

Background:

• Light field imaging captures intensity and direction, enabling immersive 6DoF experiences.
• Existing Light Field Image Quality Assessment (LFIQA) methods struggle with angular consistency and high computational costs.
• A need exists for efficient LFIQA metrics that accurately assess angular quality.

Purpose of the Study:

• To propose a novel "anglewise attention" concept for LFIQA.
• To develop efficient attention kernels for angular feature extraction.
• To introduce the Light Field Attentional Convolutional Neural Network (LFACon) as an LFIQA metric.

Main Methods:

• Introduced multihead self-attention to the angular domain of light field images.
• Proposed three new attention kernels: anglewise self-attention, anglewise grid attention, and anglewise central attention.
• Developed the LFACon metric by incorporating these attention kernels.

Main Results:

• The proposed LFACon metric significantly outperforms existing state-of-the-art LFIQA methods.
• LFACon achieves top performance across most distortion types.
• The metric demonstrates lower complexity and reduced computational time compared to prior methods.

Conclusions:

• Anglewise attention effectively enhances LFIQA by considering angular consistency.
• LFACon provides a computationally efficient and accurate solution for LFIQA.
• The proposed method advances the field of virtual reality and image quality assessment.