Image Processing Reference
In-Depth Information
Fig. 6.12 Validation framework
are too high. In the same way, the quantization step increase between adjacent
regions must be minimized in order to achieve more consistent and efficient
encoding. Different saliency to quantization offset mappings are considered and
presented in the next sections.
By using the eye tracking information, the performances of the proposed
approach are measured and compared to the reference. The validation framework
is illustrated in Fig.
6.12
. In this framework, the pre-processing module consists in
mapping the image regions with the corresponding quantization factor offsets. After
encoding, the video quality is objectively measured after masking the image with
the eye tracked areas. In this way, the measured quality corresponds to the quality of
the pixels subject to higher visual attention as measured by the eye tracking system.
In the next sections, the performances measured with the two sequences “Actor”
and “Musicians” are presented.
6.3.3.1 Disparity Map Approach
The algorithm described in Sect.
6.2.5
aims at converting disparity maps into
saliency based image partitioning where the most salient region corresponds to
the white pixels. In this approach, the saliency computation is based on a simple
heuristic (objects in the foreground are more salient than the background) and
having in mind that the saliency maps is used further to manage the bit rate
allocation at encoding stage. This second constraint leads to increase the saliency
of the pixels that belong to object borders. Finally, the Qp offsets maps introduced
previously are obtained by producing the negative image of the saliency maps as
depicted in Fig.
6.13
.
The rate-distortion curves obtained with the two sequences after H.264
encoding
6
are provided in Fig.
6.14
. In this figure, we can see the interest of the
approach, especially when the bit rate increases. For instance, for the Actor
sequence at a quality level of 42 dB, the encoding with VAM allows saving around
25% of the bit rate compared to the classical encoding approach, when measuring
