Learning to Assess Image Quality Like an Observer.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Human observers possess sophisticated visual perception abilities, integrating global and local information for image quality evaluation.
• Developing computational models that accurately predict human perception of image quality is a significant challenge in image processing.

Purpose of the Study:

• To propose a novel observer-like network (OLN) for image quality assessment (IQA) that accurately predicts human perception.
• To design a model that jointly considers global glimpsing and local scanning information, mimicking human visual behavior.

Main Methods:

• The proposed observer-like network (OLN) integrates a global distortion perception (GDP) module and a local distortion observation (LDO) module.
• The GDP module classifies image distortion categories and levels, simulating global perception.
• The LDO module simulates local observation by tracing human scanpaths, gathering long-term regional information.

Main Results:

• The OLN effectively combines global and local information using a bilinear pooling layer.
• The network accurately predicts distorted image quality scores, aligning with human observer assessments.
• Comprehensive experiments on public datasets demonstrate state-of-the-art performance.

Conclusions:

• The observer-like network (OLN) provides a powerful new approach to computational image quality assessment.
• By simulating human visual perception, the OLN achieves superior accuracy in predicting image quality.
• The model's ability to integrate global and local visual cues represents a significant advancement in IQA research.