Saliency and Power Aware Contrast Enhancement for Low OLED Power Consumption

The mass adoption of display devices calls for advanced power-reduction techniques. Power-constrained contrast enhancement (PCCE) gains many improvements in recent years, yet several important issues were neglected. Computation complexity increases significantly when processing high-resolution content, which is commonplace nowadays. Moreover, many works focus on improving a finite set of quality metrics while abandoning the salient information of image content. We propose to develop an efficient, saliency aware, on-demand PCCE method that is end-to-end optimized on image quality and saliency criterion. Our method starts by extracting multilevel features from a low-resolution luminance input using an efficient feature encoder. A lightweight power-attention mechanism realizes the on-demand power reduction via input image statistics. The last stage mitigates the artifacts introduced by the low-resolution saliency information using fast-guided filtering (GF) and local enhancement (LEN) to restore the high-frequency component. To bridge the unsupervised PCCE and supervised saliency task, we develop a local quality measure that captures a quality ratio given a desired power level. Experiments on multiple datasets with up to 4K resolution demonstrate the effectiveness of our method to produce high-quality and saliency scores. With a 20% power reduction on RAISE dataset, our method achieves structural similarity (SSIM) of 0.99 with backbone network computation of fewer than 0.1 giga multiply-accumulate operations per second (GMACs). Measurement on an organic light-emitting diode (OLED) panel indicates that our method can achieve 0.83 SSIM with a 61% reduction rate. The implementation is available at https://github.com/kuntoro-adi/SPACE.