Device and method for creating a saliency map of an image
View Patent ↗Detection of the salient points in an image enable the improvement of further steps such as coding or image indexing, watermarking, video quality estimation. The methods rely on the fact that a model is fully based on the human visual system (HVS) such as the computation of early visual features, and the methods compute a saliency map for video images taking into account motion and the velocity of the eye.
1. Method for creating a saliency map of an image comprising the step of hierarchical decomposition of the image into frequency sub-bands, comprising the steps of
movement estimation between current image and a preceding image calculated from the decomposition into frequency sub-bands,
estimating a dominant movement for the image and a local movement for each pixel of the current image,
obtaining a relative motion vector for each pixel of the image, being the difference between a local motion estimation and the dominant movement of the image of said pixel,
wherein it further comprises the steps of
if the value of a norm of the relative motion vector is higher than a maximal pursuit velocity of an eye, multiplying the relative motion vector by the maximal pursuit velocity of an eye over the norm of the relative motion vector,
computing a histogram of the module of the obtained relative motion vectors of an image in order to obtain a normalized temporal saliency map.
2. Method according to claim 1 wherein, further to the step of frequency sub-band decomposition, it comprises a step of orientation-based decomposition giving a perceptual sub-band decomposition.
3. Method according to claim 2 wherein it comprises a step of intra masking of different perceptual sub-bands.
4. Method according to claim 3 wherein it comprises:
a step of construction of a four-level pyramid,
a step of calculating a local motion vector for each pixel for each resolution level of the pyramid.
5. Method according to claim 4 wherein it comprises a step of obtaining a parametric model of the dominant movement of the image for each pixel of the image by using the local motion vectors obtained from a full resolution level of said four-level pyramid.
6. Method according to claim 1 wherein it comprises a step of computing a spatial saliency map of said image and a step of fusion of said normalized temporal saliency map and of said spatial saliency map.