Low-complexity motion-based saliency map estimation for perceptual video coding

In this paper, a low-complexity motion-based saliency map estimation method for perceptual video coding is proposed. The method employs a camera motion compensated vector map computed by means of a hierarchical motion estimation (HME) procedure and a Restricted Affine Transformation (RAT)-based modeling of the camera motion. To allow for a computationally efficient solution, the number of layers of the HME has been restricted and the potential unreliable motion vectors due to homogeneous regions have been detected and specially managed by means of a smooth block detector. Special care has been taken of the smoothness of the resulting compensated camera motion vector map to avoid unpleasant artifacts in the perceptually-coded sequence, by including a final post-processing based on morphological filtering. The proposed saliency map has been both visually and subjectively assessed showing quality improvements when used as a part of the H.264/AVC standard codec at medium-to-low bitrates.