Energy efficient algorithms and techniques for wireless mobile clients 5b

Original image Linear darkening Gamma compression HVS based Power Model
approach approach mapping based mapping
Power: 494.0 mW Power: 245.7 mW Power: 246.2 mW Power: 245.2 mW Power: 245.3 mW
GCL: 0 GCL: 0.5 GCL:12.3 GCL: 2.9 GCL: 10.0
PSNR: ∞ dB PSNR: 14.731 dB PSNR: 15.219 dB PSNR: 14.311 dB PSNR: 15.185 dB
MSSIM: 1.000 MSSIM: 0.902 MSSIM: 0.932 MSSIM: 0.902 MSSIM: 0.942
Power: 256.1 mW Power: 127.7 mW Power: 127.2 mW Power: 127.1 mW Power: 127.1 mW
GCL: 0 GCL: 0.3 GCL: 14.4 GCL: 1.7 GCL: 12.5
PSNR: ∞ dB PSNR: 16.143 dB PSNR: 17.921 dB PSNR: 16.929 dB PSNR: 17.781 dB
MSSIM: 1.000 MSSIM: 0.810 MSSIM: 0.921 MSSIM: 0.901 MSSIM: 0.928
Power: 142.2 mW Power: 71.1 mW Power: 71.1 mW Power: 71.1 mW Power: 71.1 mW
GCL: 0 GCL: 4.3 GCL: 12.9 GCL: 5.0 GCL: 12.9
PSNR: ∞ dB PSNR: 18.826 dB PSNR: 20.359 dB PSNR: 18.976 dB PSNR: 20.232 dB
MSSIM: 1.000 MSSIM: 0.664 MSSIM: 0.926 MSSIM: 0.748 MSSIM: 0.940
Figure 4.27. Power consumption and quality measurements
149
The user study shows mixed results. The results after removing biased entries are
shown in Figure 4.28. Normalised performance count is a ratio between total number
of times an image appeared in the pair for selection and number of times it is selected
as the best. For high power saving, PM based approach outperforms other approaches
and for low power saving, HVS based approach performs better. This may be due to
the capability of HVS approach to retain global contrast when minimal changes are
made. Hence, we planned to deploy a power level adaptive approach for our cloud
service.
0"
0.1"
0.2"
0.3"
0.4"
0.5"
0.6"

0.7"
20%"
40%"
60%"
80%"
Linear"Darkening"
Gamma"Compession"
HVS"Based"
Power"Model"Based"
Power"Saved"
Preference"Co unt"(Normalised)"
Figure 4.28. Image Transformation - User Study
4.7.3 Overall Result (Combined)
We now combine the results of both text and image transformations. The trade-
offs involved in saving OLED display power are when high energy saving is required,
it is clear that the key image based text/background colour transformation approach
should be combined with power-model based image transformation approach to save
150
60% (and above) power while providing best possible quality. With this combination,
the users experienced an additional 50ms to 100ms RTT delay in getting the contents
through the cloud for up to 32 users accessing the service concurrently. The RTTs
were measured while the service was run as a personal cloud service in a DELL Power
Edge T610 Tower Server with a hexa-core processor, 24GB RAM and 1 TB HDD.
For saving power up to 50% it is recommended to use HVS based approach or simple
darkening approach with the text transformations. The advantage of simple darkening
approach is in its simplicity. With simple darkening, the users have experienced less
than 50ms additional RTT delay due to lower processing latency. To support large
number of users, existing powerful clouds with cluster of servers can be used. The
default setting in our cloud service is 20% power saving, which generates content
quality (text and images) comparable to original.

4.7.4 Summary
The results show that colour transformations constrained by brand identity for
texts can save significantly more energy than all the previous works while retaining
the readability of the contents. In certain cases, it even improves the website’s colour
schemes. For images, our scheme can save the much higher level of energy while
retaining the image fidelity and keeping the distortions at minimum level when com-
pared to previous browser dependant approaches which either makes everything look
greenish [6] or saves less than 5% energy [59]. With our approaches, 60% energy can
be saved for web pages with images while providing good quality. For image free web
pages (most mobile web pages do not have images) we can save up to 80% of energy
if dark background is acceptable.
151