Blind Image Quality Assessment With Active Inference.

Area of Science:

• Computer Vision
• Image Processing
• Human Visual System Modeling

Background:

• Blind Image Quality Assessment (BIQA) is crucial but challenging.
• Mimicking the Human Visual System (HVS) offers a promising approach for BIQA.
• The Internal Generative Mechanism (IGM) highlights the HVS's active content inference for understanding.

Purpose of the Study:

• To develop a novel BIQA metric by simulating the IGM's active inference process.
• To predict an image's primary content for more accurate quality assessment.
• To improve BIQA performance, especially in cross-database evaluations.

Main Methods:

• An active inference module using Generative Adversarial Networks (GANs) to predict primary content, optimizing for semantic similarity and structural completeness.
• A multi-stream Convolutional Neural Network (CNN) to evaluate image quality based on scene information, distortion type, and content degradation, correlating with the predicted primary content.
• Quality measurement derived from the interplay between the predicted primary content and comprehensive degradation analysis.

Main Results:

• The proposed BIQA metric achieves competitive performance across five standard IQA databases.
• Significant improvements are observed in cross-database evaluations, demonstrating robustness.
• The method effectively utilizes predicted primary content and multi-stream CNN analysis for quality assessment.

Conclusions:

• The novel BIQA metric, inspired by the IGM and employing GANs and CNNs, provides an effective approach to image quality assessment.
• Mimicking the HVS's active inference process leads to superior BIQA performance.
• The method shows promise for reliable image quality evaluation in diverse scenarios.