用於反射式電子紙顯示器色彩校正的AI模型：邊緣實現的即時方法

Abstract

本研究旨在解決反射式彩色電子紙的非線性色彩失真與色偏問題。為此本研究提出了一套基於深度學習的色彩校正架構和基於過往方法開發的半色調演算法，以有效提升色彩還原的準確度與空間連續性，克服傳統方法的限制。為驗證本方法在邊緣運算裝置上的可行性與即時性，我們將模型部署於 NVIDIA Jetson Orin NX，並採用訓練後量化策略將模型由全精度轉換為INT8精度。實驗結果顯示，量化後模型在推論速度上提升近五倍，同時僅有輕微的影像品質減損，大幅降低了記憶體與運算資源需求。本研究提供了一套低成本、高效率且無需額外色彩量測的AI調色方案，證實其具備高度的實用性與延展潛力。

This study aims to address the issues of non-linear color distortion and color deviation in reflective printed color e-paper. To this end, this research proposes a deep learning-based color correction framework and a halftoning algorithm developed from previous methods to effectively enhance the accuracy of color reproduction and spatial continuity, overcoming the limitations of traditional approaches. To validate the method's feasibility and real-time performance on edge computing devices, we deployed the model on an NVIDIA Jetson Orin NX and employed a post-training quantization strategy to convert the model from full precision to INT8 precision. Experimental results demonstrate that the quantized model achieved a nearly five-fold increase in inference speed with only a slight degradation in image quality, significantly reducing memory and computational resource requirements. This research provides a low-cost, high-efficiency, and measurement-free AI color tuning solution.