使用深度學習從單張低動態範圍影像重建高動態範圍影像

Abstract

我們提出了一個基於深度學習的方式，從單張低動態範圍影 像 (LDR) 重建高動態範圍影像 (HDR) 。這個問題非常具有挑戰性， 因為目前相機結構上的限制，在拍攝照片時會有影像量化的問題 (quantization) 以及過曝的問題。與前人所提出的方法不同點是:我們提出的網路架構並不直接學習低動態範圍影像與高動態範圍影像之間的對應關係，我們在網路架構中加入了更多人類關於相機與影像的知識。明確的說，我們提出的網路架構含有三個部分:(1) 影像線性化 (2) 影像去量化 (3) 影像去過曝。第一，給訂一張低動態範圍影像，我們先找出對應的相機反應函數，便可以將影像轉換到線性的空間中。 第二，我們透過影像去量化網路來去除影像過暗區域所造成的影像量化問題。第三，我們使用影像去過曝網路來重新繪畫出因為過曝所缺失的影像內容。最後，我們對所提出的演算法與同類型演算法做了許多實驗與比較，證明我們的演算法優於前人所提出的演算法。

We present a learning-based approach for recovering a high dynamic range (HDR) image from a single low dynamic range (LDR) input image. This problem is challenging due to missing details in under-/over-exposed regions caused by quantization and saturation of camera sensors. In contrast to ex- isting methods that directly learn the mapping from LDR to HDR images using a generic network, we propose to integrate the domain knowledge of the LDR image formation pipeline into our model. Specifically, our ap- proach consists of three modules to restore HDR details from LDR images: (1) linearization, (2) dequantization, and (3) hallucination. First, given an input LDR image, we estimate the inverse camera response function using a Linearization-Net and map the LDR image to the linear space. Second, we apply a Dequantization-Net to remove quantization artifacts in the under- exposed regions. Third, we inpaint the missing regions due to saturation with a Hallucination-Net. Extensive quantitative and qualitative experiments demonstrate that our approach performs favorably against state-of-the-art tech- niques by learning to reverse the image formation pipeline.