Luminance compensation for stretchable displays using deep visual feature-optimized Gaussian-weighted kernels

Abstract

Stretchable displays, characterized by their flexibility and deformability, are gaining attention as next-generation display technologies. While various studies have been conducted on hardware aspects of stretchable displays, the software aspects have received comparatively less focus. When displays are stretched, empty pixels inevitably lead to a decrease in overall luminance, which significantly degrades visual quality and user experience. To address this issue from a software aspect, we propose a novel luminance compensation method that leverages deep learning through a Learned Perceptual Image Patch Similarity (LPIPS)-based pre-optimization technique combined with Gaussian-weighted kernels. The proposed method applies relatively higher values to areas near empty pixels, where luminance loss is most significant while preserving the original luminance in unaffected areas. This design minimizes color distortion and enhances brightness effectively. Specifically, the optimal brightness increase rates (BIRs) are pre-optimized using an LPIPS-based loss function, tailored to various stretching scenarios, such as stretching types, directions, and rates. Based on the optimized BIRs, Gaussian-weighted kernels are generated for efficient luminance adjustment. Our method flexibly supports diverse stretching conditions, including linear/non-linear stretching and uni-directional/bi-directional stretching, with stretching ratios ranging from 10% to 30%. Through simulations, we qualitatively and quantitatively compared the proposed method with existing approaches, demonstrating superior performance across a wide range of scenarios.