基於擴散模型的高動態範圍圖片之重建

Abstract

HDR 重建任務傳統上是通過使用多張不同曝光的 LDR 影像來完成的。本文提出了一種新的方法，首先使用隱式函數模型從單一曝光的 LDR 生成多曝光的LDR 堆疊。然後，將 HDR 重建視為一個影像生成任務，其中 LDR 堆疊的特徵被用作擴散模型的輸入條件來生成 HDR 影像。 與直接將條件輸入擴散模型不同，我們選擇使用帶有注意力機制的條件特徵生成器來引導基於參考影像的特徵融合。為了解決監督式擴散模型中的顏色失真問題，加入了直方圖損失（Histogram Loss），有效地校正了生成影像中的顏色偏移。實驗結果顯示，這種方法在評估指標和真實世界影像質量上表現良好，並具有較強的泛化能力。

The HDR reconstruction task is traditionally accomplished using multiple LDR images with different exposures. This paper introduces a new approach that begins by using an implicit function model to generate a multi-exposure LDR stack from a middle exposure (EV0) LDR image. Then, HDR reconstruction is treated as an image generation task, where the features of the LDR stack are used as input conditions for a diffusion model to create the HDR image. Instead of directly inputting the condition into the diffusion model, we choose to use a Conditional Feature Generator with an attention mechanism to guide the merging process based on the reference image. To solve the issue of color distortion in supervised diffusion models, a Histogram Loss is added, which effectively corrects color shifts in the generated images. Experimental results show that this method performs well in both evaluation metrics and real-world image quality, with strong generalization ability.