Perception-Inspired Network for Stereo Image Quality Assessment

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Image Processing

Background:

• Current stereo image quality assessment (SIQA) methods struggle with accurate modeling of binocular fusion and fine-grained perception.
• Limitations exist in capturing the nuances of human visual system (HVS) processing for stereo images.

Purpose of the Study:

• To develop an advanced SIQA method that simulates key mechanisms of the HVS for more accurate quality prediction.
• To overcome the limitations of existing SIQA models in binocular fusion and detailed perception.

Main Methods:

• A Perception-Inspired Network for SIQA was proposed, incorporating a difference-guided binocular fusion (DGBF) module.
• A high-frequency enhancement module (HFEM) was developed to simulate HVS sensitivity to distortions in high-frequency components.
• A hierarchical quality regression strategy was implemented to mimic human perception from local details to global judgment.

Main Results:

• The proposed method demonstrated superior performance compared to mainstream SIQA approaches.
• Achieved a high Pearson Linear Correlation Coefficient (PLCC) of 0.9734 on the LIVE I database.
• Achieved a PLCC of 0.9632 on the LIVE II database, indicating strong alignment with subjective evaluations.

Conclusions:

• The novel Perception-Inspired Network effectively simulates HVS mechanisms for improved SIQA.
• The DGBF and HFEM modules, along with hierarchical regression, enhance the accuracy of stereo image quality prediction.
• The proposed method offers a more reliable and human-aligned approach to assessing stereo image quality.