Unsupervised Deep Contrast Enhancement With Power Constraint for OLED Displays

Yong-Goo Shin, Seung Park, Yoon-Jae Yeo, Min-Jae Yoo, Sung-Jea Ko

Abstract

Various power-constrained contrast enhancement (PCCE) techniques have been applied to an organic light emitting diode (OLED) display for reducing the power demands of the display while preserving the image quality. In this paper, we propose a new deep learning-based PCCE scheme that constrains the power consumption of the OLED displays while enhancing the contrast of the displayed image. In the proposed method, the power consumption is constrained by simply reducing the brightness a certain ratio, whereas the perceived visual quality is preserved as much as possible by enhancing the contrast of the image using a convolutional neural network (CNN). Furthermore, our CNN can learn the PCCE technique without a reference image by unsupervised learning. Experimental results show that the proposed method is superior to conventional ones in terms of image quality assessment metrics such as a visual saliency-induced index (VSI) and a measure of enhancement (EME).

Introduction

OLED displays are inherently power-intensive because each pixel acts as an independent light source. Consequently, higher brightness levels directly correlate with increased energy consumption. While dimming the display is a straightforward way to save power, it often leads to a significant loss in image contrast and overall visibility. Existing Power-Constrained Contrast Enhancement (PCCE) techniques attempt to mitigate this by dimming the image and then applying mathematical filters to restore contrast. However, traditional algorithms frequently introduce unpredictable artifacts or fail to preserve fine details.

Applying deep learning to PCCE is challenging because there is no "ground truth" reference for what an optimally power-saved image should look like. This paper addresses these issues by introducing an unsupervised deep learning framework that reduces brightness to meet power targets while simultaneously boosting contrast to maintain perceived quality.

Proposed Method: The ACE Network

The core of the proposed system is the Adaptive Contrast Enhancement (ACE) network, which is based on a modified Context Aggregation Network (CAN). The system follows a pipeline where an input luminance image is darkened by a factor $k$ to meet a power-saving target, processed by the ACE network for enhancement, and then recombined with the original color information.

Overview of the proposed power-constrained contrast enhancement (PCCE) system using the ACE network.

Unsupervised Learning and Loss Functions

To overcome the lack of reference images, the authors utilize a multi-objective loss function consisting of three components:

  1. Power Loss ($L_p$): Ensures the output adheres to a specific power-saving target ($R$).
  2. Similarity Loss ($L_s$): Employs the Structural Similarity Index (SSIM) to prevent visual distortions relative to the original image.
  3. Contrast Loss ($L_c$): A custom loss that encourages higher standard deviation in pixel values globally and locally to maximize visual "pop."

Architectural Innovations

The ACE network utilizes dilated convolutions to maintain a lightweight profile suitable for hardware implementation. Dilated filters allow the network to capture a larger global context without increasing parameter counts. Additionally, the network accepts the desired power-saving rate ($R$) as a conditional input, allowing a single trained model to adapt to various energy-saving requirements (e.g., 10% vs. 70%).

Experimental Results

The model was trained on the Berkeley Segmentation Dataset (BSD) and validated using the LIVE dataset and cinematic video sequences from The Shawshank Redemption and Avatar.

Quantitative and Qualitative Performance

The ACE network consistently outperformed state-of-the-art traditional algorithms, such as HPCCE and LAPSE, across key metrics:

  • Metrics: Superior scores in the Measure of Enhancement (EME) and Visual Saliency-induced Index (VSI) indicate better contrast and perceived quality.
  • Efficiency: The model processed 720p images in approximately 31.2 ms on a Titan X GPU, demonstrating potential for real-time video applications.
  • Visual Fidelity: Qualitative analysis shows that the ACE network avoids the "crushed" blacks and halo artifacts common in other methods, preserving fine textures like fur and foliage even at high power-saving rates.
Qualitative comparison of the proposed method against existing techniques, highlighting detail preservation in zoomed patches.

Conclusion and Future Work

By leveraging unsupervised learning, the ACE network discovers enhancement strategies that traditional mathematical models might overlook. While the current architecture is compact, future research will focus on further reducing computational complexity to ensure that the energy saved by dimming the OLED panel is not offset by the power required to run the AI on mobile processors or FPGA-based accelerators.

Summary updated

Abstract

Various power-constrained contrast enhance-ment (PCCE) techniques have been applied to an organic light emitting diode (OLED) display for reducing the pow-er demands of the display while preserving the image qual-ity. In this paper, we propose a new deep learning-based PCCE scheme that constrains the power consumption of the OLED displays while enhancing the contrast of the displayed image. In the proposed method, the power con-sumption is constrained by simply reducing the brightness a certain ratio, whereas the perceived visual quality is pre-served as much as possible by enhancing the contrast of the image using a convolutional neural network (CNN). Furthermore, our CNN can learn the PCCE technique without a reference image by unsupervised learning. Ex-perimental results show that the proposed method is supe-rior to conventional ones in terms of image quality assess-ment metrics such as a visual saliency-induced index (VSI) and a measure of enhancement (EME).1.
BibTeX
@article{shin2019unsupervised,
  author = {Yong-Goo Shin and Seung Park and Yoon-Jae Yeo and Min-Jae Yoo and Sung-Jea Ko},
  title = {{Unsupervised Deep Contrast Enhancement With Power Constraint for OLED Displays}},
  journal = {IEEE Transactions on Image Processing},
  year = {2019},
  doi = {10.1109/tip.2019.2953352},
}
Figure