Abstract
Due to the popularity of online repositories, an efficient compression algorithm removes not only statistical redundancy but also the psychovisual redundancies without perceptual degradation is important for transmission and storage. Visual attention model or visual sensitivity model are proposed for removing psychovisual redundancies. Most visual attention models are based on spatial component analysis and only few adopt motion vectors for temporal component analysis which lacks of perceptual information. Scene motion is a powerful feature for identification of regions of interest (ROIs) since it indicates the producer’s interests in the scene and can also attract the viewer’s attention. In addition to the global scene motion, the motion of salient object in the video stream facilitates tracing salient object locally. In this paper, we propose a scene motion and saliency motion based visual attention model to effectively trace the movement of salient regions. We propose a frame that incorporates the obtained motion saliency map with Just Notice Difference (JND) as a visual measure to determine the quantization parameters at the macro-block level. To evaluate the performance of our proposed framework, three experiments have been executed for verifying the accuracy of saliency motion prediction, the video compression rate and the visual quality assessment. Experiment results show our proposed framework has higher saliency prediction accuracy than previous approaches in term of Receiver Operating Characteristics (ROC). Our proposed framework achieves 8% up to 73% bit rate reduction compared with the H.264 in version J14.0 and the bit rate reduction is three times higher compared with the previous method. From the visual quality assessment experiments, participants cannot distinguish the difference between our compressed video and the original video streams.