Pyramidal modeling of geometric distortions for retargeted image quality evaluation

Content-aware retargeting methods are used to adjust images to different resolutions and aspect ratios with low deformation and information loss in salient regions. Effective objective quality assessment of retargeted images can provide a way to improve retargeting methods. The non-uniform geometrical degradations caused by retargeting algorithms make it impossible to use traditional image quality assessment metrics for retargeted images. Although some quality evaluation methods have been proposed till now, the resulted quality scores are not well correlated with the subjective ones. In this paper we propose a pyramidal global-to-local pooling method to combine pixel/block deformation measures. In each level of locality, the Euclidean distance between the retargeted and original image is used as an individual feature. Therefore, in addition to the global summation, assessment of local deformations, contributes toward better quality evaluation. Learning a regression model based on the extracted features results in better performance compared to relevant existing retargeted image quality methods.