Energy-aware Images: Reducing the energy consumption of OLED displays - Extended abstract

Abstract

In the context of climate change, it is necessary to reduce as much as possible the energy consumption in the video chain. In this demonstration, we show how to reduce the consumption of displaying images on OLED screens, thanks to a shallow network specifically trained to build energy-aware images. We demonstrate qualitative and quantitative results through several metrics and real measures of energy consumption. The performances are assessed against other methods in the state-of-the-art.

Keywords

• eol>energy consumption
• display devices
• energy-aware images
• attenuation map

1. Introduction

2. Deep PVR

Our proposed model, called deepPVR for Deep Pixel Value Reduction, is illustrated in Figure 1. It is inpired from the R-ACE network proposed in [ 3 ], but with a revisited shallow network architecture, to reduce as much as possible the number of trainable parameters. Indeed, in accordance with our willingness to globally reduce the energy consumption, one also needs to pay a special attention to the size of the proposed networks, which directly participates to the needed training duration.

To this aim, the number of channels per layer was decreased and the CAN (Context Aggregation Network) was replaced by ATrous spatial pyramid pooling [ 4 ]. To improve the performance, and similarly to what was proposed in [ 5 ], a channel and spatial attention layers were added, however in a simplified version, by considering an adaptive 2D average pooling followed by two convolution layers, with respectively a ReLu and a sigmoid activations. The last two convolution layers output a dimming map which is simply added to the luminance of the input image to generate the energy-aware image. All convolution layers have a 3 × 3 kernel. A weighted average of the MAE loss, the SSIM loss, a power loss and a Total Variation (TV) loss was used as training objective. The power loss is computed as the diference between the power of the reduced image and the power of the original image reduced by a factor 1 − , where is the target energy reduction factor. The TV loss is computed on the dimming map as the average of the squares of its vertical and horizontal gradients.

DeepPVR architecture totals less than 5k parameters, which is an order of magnitude lower than the 41k parameters needed by R-ACE network, and far less than the 1.8 million parameters in [ 6 ].

3. Comparison with other methods

To illustrate the performances of the proposed algorithm, we first compare with a simple linear scaling, which consists in scaling down evenly the luminance of some input image. For this purpose, we determine a scaling coeficient to reach an energy consumption target , such ^ , where ^ and represent the powers dissipated by the screen when that: = 1 − deepPVR R-ACE LS ori

10% displaying the processed and the original images, respectively. Assuming that the energy model is linear, the scaling coeficient is given by: = (1 − )1/ , where is the gamma correction of the screen.

We also compare with our own implementation of the R-ACE network [ 3 ]. In particular, we trained R-ACE network on the BSDS dataset [ 7 ] with the same loss functions used for deepPVR.

4. Evaluation

In the proposed demonstration, we illustrate the performances of our approach through results on images while comparing with the three methods described in section 3. An example of such results is shown in Figure 2.

A quantitative evaluation (see Table 1) is also discussed in details. Additional real measures of energy consumption are also demonstrated, together with power consumption maps.

Acknowledgments

This work has been achieved in the context of the project 3EMS-2 funded by the “Région Bretagne”, Rennes Métropole, co-funded by E.U and supported by “Images et Réseaux. (a) Original (b) DeepPVR

References

[1] T. C. Trust , Carbon impact of video streaming , https://www.carbontrust.com/en-eu/node/1537, 2021 .

[2] O. Le Meur , C.-H. Demarty , E. Reinhard , F. Aumont , L. Blondé, Energy-aware images: Quality of experience vs energy reduction, in: 2023 ACM Mile-High Video (MHV) , ACM, 2023 .

[3] K. A. Nugroho , S.-J. Ruan , R-ace network for oled image power saving , in: 2022 IEEE 4th Global Conference on Life Sciences and Technologies (LifeTech) , IEEE, 2022 , pp. 284 - 285 .

[4] L.-C. Chen , G. Papandreou , F. Schrof , H. Adam , Rethinking atrous convolution for semantic image segmentation , arXiv preprint arXiv:1706.05587 ( 2017 ).

[5] J. Park , S. Woo , J.-Y. Lee , I. S. Kweon , Bam: Bottleneck attention module, arXiv preprint arXiv: 1807 . 06514 ( 2018 ).

[6] J.-L. Yin , B.-H. Chen , Y.-T. Peng, C.-C. Tsai, Deep battery saver: End-to-end learning for power constrained contrast enhancement , IEEE Transactions on Multimedia 23 ( 2020 ) 1049 - 1059 .

[7] D. Martin , C. Fowlkes , D. Tal , J. Malik , A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics , in: Proc. 8th Int'l Conf . Computer Vision , volume 2 , 2001 , pp. 416 - 423 .

[8] F. R. , Kodak lossless true color images , 1999 . URL: https://r0k.us/graphics/kodak/.