Image Processing Reference
In-Depth Information
bit rate compared to the classical encoding approach. As for the disparity based
approach, this bit rate saving is obtained by decreasing the quality of the less salient
regions which are subject to lower perceptual attention. However, this quality
decrease remains limited and can be adjusted following the application require-
ments. In order to adjust the level of differentiation between regions, the quantiza-
tion is a key process. For instance, computing the quantized maps as follows:
S
q
ΒΌ
b
S
1
=
4
c
will merge more dramatically the salient regions and decrease the
maximum Qp offset from 6 to 4. In terms of performances, this would reduce the
quality gain on the salient pixels and better preserve the less salient areas.
Conclusion
In this chapter, we have firstly listed the main approaches, classically used to
extract saliency information from 2D video signals. The more efficient
approaches consist in processing jointly different clues of the signal: spatial
and temporal information. In addition, the content considered in this chapter
is multi-view and third dimensional data are freely available in order to build
a quick and efficient saliency computation algorithm based on disparity maps.
In addition to this fast and basic implementation, an improved version of the
temporal image signature based approach is proposed.
In order to assess the proposed saliency map extraction methods, an eye
tracking system has been used to produce reference material and to compare
the algorithms outputs with the eye tracker data. Statistically, we have
observed a good matching between the saliency maps produced by the TIS
method and the eye tracking data.
Moreover, a specific use case where saliency is used to improve bit rate
allocation in a video encoder has been tested. With the disparity based
approach, we can easily, without requiring high computation resources,
improve the rate-distortion performances when focusing on pixels referenced
by the eye tracking tests. With the improved TIS approach, the performances
gains on the salient pixels is even better and especially at high bit rates, the
approach achieves quite impressive results.
Finally, this study confirms the interest and benefit of visual attention
modelling when designing advanced image processing applications (e.g.,
3D multi-view video) and the integration of these approaches in complete
application frameworks is really promising when considering QoE optimiza-
tion aspects.
Acknowledgements This work was supported by the ROMEO project (grant number: 287896),
which was funded by the EC FP7 ICT collaborative research program.
