Image retargeting quality assessment based on saliency-driven classification

Abstract

The overall geometric distortions of the visually prominent contents in resized images are sensitive to human perception. But the existing image retargeting quality assessment (IRQA) algorithms did not consider it, leading to a lower performance of evaluation algorithms. Additionally, the visually important region needs to be extracted from the retargeted images to measure their distortions. However, due to the size change and content alternation during retargeting, the visually prominent regions of the resized images obtained by the existing saliency detection approaches do not match the corresponding portions of the original photos. To address these issues, we propose an IRQA method based on saliency-driven classification, which uses various metrics to evaluate the qualities of different retargeted images. We also design a novel algorithm to extract the salient region of the retargeted image, which utilizes the original image to perform backward registration and position matching of the salient area. Besides, we present an improved aspect ratio similarity (ARS) metric to measure the overall geometric distortion occurring in the visually important region of the resized images. Experimental results show that the performance of the proposed IRQA method is superior to the other existing algorithms.