SNIC: Synthesized Noisy Images using Calibration
Nik Bhatt
TL;DR
This paper created the Synthesized Noisy Images using Calibration dataset (SNIC), containing over 6000 noisy images, comprising 30 scenes from four sensors, including two smartphone sensors, a point-and-shoot, and a DSLR, the first synthesized noisy image dataset provided in both RAW and TIFF format.
Abstract
Advanced denoising algorithms require large, high-quality datasets. Physically-based statistical noise models can create such datasets by realistically simulating noise in digital images. However, there is little information on the correct way to calibrate and tune these heteroscedastic models, and a lack of published datasets using them. In this paper, we explore the process of building high-quality heteroscedastic noise models. Our methods produce realistic synthesized noisy images in both RAW and TIFF formats. Our synthesized noisy images achieve comparable LPIPS results to real noisy images; when tested with a state-of-the-art denoising model, our images reduce the PSNR gap versus real noise by 54-64% compared to those synthesized using manufacturer-provided DNG noise models. Using our approach, we created the Synthesized Noisy Images using Calibration dataset (SNIC) containing over 6000 noisy images, comprising 30 scenes from four sensors, including two smartphone sensors, a point-and-shoot, and a DSLR. SNIC is the first synthesized noisy image dataset provided in both RAW and TIFF format.