Color face-tuned salient detection for image quality assessment

Yubing Tong (1) , Hubert Konik (1) , Alain Trémeau (1)

1 LHC - Laboratoire Hubert Curien

PSNRHVS and PSNRHVSM are two new emerging image quality assessment methods but they fail when assessing the quality of some distorted images called as "extreme" images. In this paper an algorithm is proposed to enhance their performance on extreme images while keeping their good performance on "normal" images unchanged. First, extreme images derived from PSNRHVS are labeled with an iterative algorithm. Then an SVM classifier is used to decide if current images are extreme images or not. Next, region saliency information is computed only for this kind of images. Then region saliency information is used instead of point saliency information in image quality assessment. We use color, intensity and orientation to compute the saliency of regions. We use also a face descriptor as faces play an important role in visual perception. The algorithm that we propose has been tested on the TID2008 database. The results that we have obtained show that the performance on extreme images is greatly enhanced compared with the original PSNRHVS.

Domaines: Traitement des images [eess.IV]

Yubing Tong, Hubert Konik, Alain Trémeau. Color face-tuned salient detection for image quality assessment. 2nd European Workshop on Visual Information Processing (EUVIP), Jul 2010, Paris, France. pp.253-260, ⟨10.1109/EUVIP.2010.5699149⟩. ⟨hal-00590690⟩