Perception-Oriented Bidirectional Attention Network for Image Super-Resolution Quality Assessment

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Super-resolution (SR) algorithms aim to enhance image resolution.
• Existing full-reference (FR) image quality assessment (IQA) metrics are limited for evaluating SR algorithms.

Purpose of the Study:

• To propose a novel FR-IQA metric for SR algorithms.
• To develop a network that accurately assesses perceived quality of SR images.

Main Methods:

• Introduced the Perception-oriented Bidirectional Attention Network (PBAN) with three modules: image encoder, perception-oriented bidirectional attention (PBA), and quality prediction.
• Constructed PBA module inspired by human visual system characteristics, incorporating Bidirectional Attention.
• Utilized Grouped Multi-scale Deformable Convolution and Sub-information Excitation Convolution for adaptive distortion perception.

Main Results:

• PBAN effectively encodes images for feature representation.
• The PBA module bidirectionally constructs visual attention to distortion.
• Experiments show PBAN outperforms existing state-of-the-art quality assessment methods.

Conclusions:

• PBAN provides a robust solution for SR image FR-IQA.
• The proposed method aligns quality assessment with human perception of distortions.
• PBAN demonstrates superior performance in evaluating SR algorithm outputs.