Brummer, Benoit
[UCL]
De Vleeschouwer, Christophe
[UCL]
Convolutional neural networks have been the focus of re-search aiming to solve image denoising problems, but theirperformance remains unsatisfactory for most applications.These networks are trained with synthetic noise distribu-tions that do not accurately reflect the noise captured byimage sensors. Some datasets of clean-noisy image pairshave been introduced but they are usually meant for bench-marking or specific applications. We introduce the NaturalImage Noise Dataset (NIND), a dataset of DSLR-like im-ages with varying levels of ISO noise which is large enoughto train models for blind denoising over a wide range ofnoise. We demonstrate a denoising model trained with theNIND and show that it significantly outperforms BM3D onISO noise from unseen images, even when generalizing toimages from a different type of camera. The Natural ImageNoise Dataset is published on Wikimedia Commons suchthat it remains open for curation and contributions. We ex-pect that this dataset will prove useful for future image de-noising applications.
Bibliographic reference |
Brummer, Benoit ; De Vleeschouwer, Christophe. Natural image noise dataset.IEEE Conference on Computer Vision and Pattern Recognition - NTIRE Workshop (Long Beach, du 16/06/2019 au 20/09/2019). In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Vol. 1, no. 1, p. 1-10 (2019) |
Permanent URL |
http://hdl.handle.net/2078.1/219128 |