Image retargeting quality assessment based on support vector regression

• Formulated major quality factors (in both spatial and frequency domains) into IRQA.
• Identified shape distortion and proposed a measure for it.
• Proposed frequency domain estimator to alleviate the measurement error in spatial domain.
• Used machine learning to integrate different quality factors.

Image retargeting is to re-represent the image, usually with a different image size to cater for the display size of the end-user (e.g., computer, smartphone). In this paper, we propose an image retargeting quality assessment scheme based on four quality factors and support vector regression. The accounted quality factors are of two categorizes: 1) shape distortions, they are visual unpleasant artifacts introduced by retargeting, e.g., a straight line may be retargeted as a curve, a circle may be retargeted as an ellipse, etc.; 2) visual content changes, they mainly refer to the loss of visual information due to retargeting. Both spatial and frequency information are used toward a complete assessment. The spatial quality factors measure shape distortions and visual content changes separately and independently, and therefore it is possible to control their relative importance to the overall assessment. Since spatial quality factors need image matching which may cause measurement error, another quality factor in frequency domain is also used. The overall image retargeting quality is contributed by the four quality measures collectively and machine learning is used accounts for such a combination, corresponding discussions are also provided for each factor. The effectiveness of the proposed scheme is verified with the publicly-available subject-rated database (comprising of diverse images and retargeting methods) and better performance (in terms of both accuracy and complexity) is achieved in comparison with the state of the art schemes.