L ’ Université de Nantes Polytech Nantes MASTER MDM “ Multimedia and Database Management

The visual attention complexity of a scene is regarded as a feature describing many information of the content in a video sequence, which meanwhile plays an important role in various applications to estimate the perceptual quality, information retrieval and so on. In this paper, the hypothesis about visual attention complexity is proposed. The complex video sequence in the view of visual attention should contain a large quantity of “informative” objects on the scene. So, the visual attention complexity(VAC) indicator extraction from a video sequence is conducted by information theory on saliency map generated from computational visual attention model. The VAC indicator’s performance is analyzed with the ground truth from an eyetrack database of IRCCyN/IVC. The proposed VAC indicator is applied in video quality estimation methods which is widely used in video transmission or compression system. In addition to VAC indicator, spatial and temporal information from original video sequence and Bitrate or PSNR from compression video compose the set of elements for the quality estimation. The objective video quality estimation model is based on a machine learning algorithm and tested on a H.264 compressed video database of IRCCyN/IVC. All proposed models are verified by another H.264 compressed video database of IRCCyN/IVC. The quality scores predicted by proposed models have a high correlation coefficient to the subjective quality scores. 1 ha l-0 08 61 08 2, v er si on 1 11 S ep 2 01 3