Quality Assessment of Retargeted Images Using Hand-Crafted and Deep-Learned Features

Since the goal of image retargeting is to adapt source images on target displays with different sizes and aspect ratios, how to objectively evaluate the quality of retargeted images is particularly important to optimize the retargeting operations. In this paper, we proposed a new image retargeting quality assessment metric, which constructs the metric using both hand-crafted features and deep-learned features. To enhance the reliability and accuracy of the proposed method: 1) we use similarity transformation as local descriptor to extract hand-craft features, and measure structure distortion and content loss from the hand-craft features and 2) we use deep learning architecture to construct encoders and extract deep-learned features, and measure texture similarity and semantic similarity from the deep-learned features. We conduct experiments on two databases: RetargetMe and CUHK. Experimental results show that our method can achieve superior performance to the state-of-the-art metrics.