Image Quality Assessment Based on Local Linear Information and Distortion-Specific Compensation.

Area of Science:

• Computer Vision
• Image Processing
• Perceptual Modeling

Background:

• Image quality assessment (IQA) is crucial in computer vision.
• Current IQA methods often fail to fully exploit distortion-specific properties correlated with human perception.
• Existing objective IQA metrics may not accurately reflect subjective human visual experience.

Purpose of the Study:

• To develop a novel IQA method that better aligns with human visual perception.
• To address the challenge of varying image distortion types in IQA.
• To improve the accuracy and robustness of objective image quality evaluation.

Main Methods:

• A local linear model is employed to analyze distortions between reference and distorted images.
• A distortion-specific compensation strategy is proposed, learning score offsets from known distortion types.
• A convolutional neural network (CNN)-based method is developed for automatic score offset computation for unknown distortions.

Main Results:

• The local linear model effectively models human perception, particularly for individual distortions.
• The proposed integrated IQA metric, combining the local linear model and compensation strategy, outperforms state-of-the-art IQA methods.
• Experimental results validate the utility of distortion-specific compensation in IQA.

Conclusions:

• The novel IQA method demonstrates superior performance in objective image quality evaluation.
• The approach effectively models human perception by considering distortion-specific properties.
• This work advances the field of image quality assessment by providing a more accurate and robust metric.