DND(Darmstadt Noise Dataset)

达姆施塔特噪声数据集
由于缺乏真实的地面实况数据，图像去噪技术传统上是在被合成 i 破坏的图像上进行评估的。一世。 d。高斯噪声。这是很成问题的，因为真实照片中的噪点不是 i。一世。 d。合成噪声过程的高斯甚至看似次要的细节，例如噪声值是否四舍五入到整数，都会对方法的相对性能产生重大影响。
因此，我们提出了一种新颖的去噪基准，即达姆施塔特噪声数据集 (DND)。它由 50 对真实噪声图像和相应的地面实况图像组成，这些图像是用不同传感器尺寸的消费级相机捕获的。对于每一对，使用基本 ISO 级别拍摄参考图像，而使用较高 ISO 和适当调整的曝光时间拍摄嘈杂图像。参考图像经过仔细的后处理，需要进行小的相机移位调整、线性强度缩放和去除低频偏差。后处理的图像作为我们去噪基准的基本事实。

Darmstadt Noise Dataset (DND)
Due to the lack of real ground-truth data, image denoising techniques have traditionally been evaluated on images corrupted by synthetic independent and identically distributed (i.i.d.) Gaussian noise. This is highly problematic, since noise in real-world photographs is not i.i.d., and even seemingly minor details such as whether noise values are rounded to integers can significantly impact the relative performance of different methods.

We therefore propose a novel denoising benchmark: the Darmstadt Noise Dataset (DND). It consists of 50 pairs of real noisy images and corresponding ground-truth images, captured using consumer-grade cameras with different sensor sizes. For each pair, a reference image is captured at the base ISO level, while a noisy image is captured at a higher ISO with appropriately adjusted exposure time. The reference images undergo careful post-processing, which includes minor camera shift alignment, linear intensity scaling, and removal of low-frequency bias. The post-processed images serve as the ground truth for our denoising benchmark.

搜集汇总

背景与挑战

背景概述

DND（达姆施塔特噪声数据集）是一个用于图像去噪的基准数据集，包含50对真实噪声图像和对应的地面实况图像，这些图像由不同传感器尺寸的消费级相机拍摄，旨在解决传统去噪评估中合成噪声与真实噪声不匹配的问题。数据集提供了经过后处理的参考图像作为地面实况，以支持更准确的去噪算法评估，由达姆施塔特工业大学于2017年发布。