|
Discriminant Saliency is a method for determining visually important regions within a scene. This is accomplished by modelling the center-surround spatial encoding nature of the human retina. For each region, a set of features is used to analyze a small window surrounding the region (center) and a larger surrounding area (surround). Saliency is measured as the ability of the feature to discriminate between the center and surround. In a collaboration with the Statistical Visual Computing Laboratory (SVCL), the dynamic texture model is used as a feature for video applications.
|
|
Discriminant Saliency has been particularly useful in increasing the performance of motion estimation for Frame Rate Up Conversion (FRUC) applciations. In this research, Motion Estimation (ME) and Motion Compensated Frame Interpolation (MCFI) are utilized in order to create interpolated frames, thereby increasing video frame rate. This technique is widely used for LCD displays and mobile video. In order to improve visual quality of FRUC, we employ motion vector refinement for salient scene regions, while enforcing smoothness of the motion field for non-salient regions.
|
