Integration of local and global features for image retargeting quality assessment

This paper presents a new objective retargeting quality assessment method based on the integration of global and local features by a Gaussian regression model. In the proposed method, first, a graph-based visual saliency algorithm and deep learning model are employed to extract an importance map of the input image. At the same time, the SIFT-Flow method is used for estimating the displacement of different pixels in the retargeted image. Then, the importance map and the displacement results are used for computing four features, called local aspect ratio similarity, local geometric distortion, global geometric distortion, and global preserved information, from the images. Finally, a Gaussian process regression model is employed to integrate these features and compute the final criterion for retargeting quality assessment. The proposed method was tested on the images of the MIT-RetargetMe and CUHK datasets and the results demonstrated its excellent performance compared to state-of-the-art methods.