李月光: 基於機器學習之1080p月光即時視訊降雜訊

Abstract

本論文提出深度學習的模型來進行低光源(月光, 0.5lux)連續影像1080p的降雜訊處理。近幾年來大眾對於手機攝影的需求以及要求有越來越大與越嚴格的趨勢。於是對於手機廠商，它們決定對於影像處理上面需要有更新穎的處理方式。除了傳統的處理方式，目前廠商的研究主要專注在深度學習對於Raw影像的生成的幫助，尤其是對於低光源目標的降雜訊以及高動態範圍成像 (HDRI: High Dynamic Range Image)。透過感光元件所獲取的最原始資訊來訓練模型。可以大幅提升深度學習的成功及成效。於是我們決定設計一套降雜訊流程，參考幾種最先進模型的優點後，提出一個新的模型，對於低光度的連續影像，提取出有影像裡面具有光照度意義的特徵點，利用其資訊對於原始影像來產生去雜訊後的擬合影像。

We propose a deep learning model targeted for low light 1080p video, which is under 0.5 lux of illumination, real-time video enhancement, and denoising. The cellphone photography market demand and competition become larger and harsher. Therefore, cellphone manufacturers decide to have novel processing algorithms in the image processing pipeline. Besides traditionally, the major market effort is dedicated to better generation of raw images using deep learning methods, especially for denoising and High Dynamic Range Image (HDRI) for low-light targets. Using raw information as the input of the model can increase the effectiveness and the success rate when training model. Thus, we design a model for a series of continuous low illumination raw image data, extract features with significant luminance features and output the denoised, enhanced image.