• 1987 (Vol.1)

Hardware independence and accuracy of neural network denoising of images depending on training set size

© 2021 N. I. Popov, A. S. Grigoryev

Moscow Institute of Physics and Technology, 171401 Dolgoprudny, Institutskiy Pereulok, 9, Russia
Institute for Information Transmission Problems, 127051 Moscow, Bolshoy Karetny per., 19, Russia

Received 02 Nov 2020

This work is an investigation of image enhancement (increasing exposure and noise reduction) using neural networks proposed in Chen et al. Learning to See in the Dark (2018). The applicability of a neural network trained on the dataset from a camera to images from another camera and possibility to reduce the size of dataset to train neural network for a new camera are analyzed. A dataset of 27 aligned pairs of raw photos was collected with Exposures of 0.01s and 1s. The model trained on dataset of 320 scenes is compared with the model trained on randomly sampled 40 scenes by testing on photos from multiple cameras with the same type of colour filter.Considering the dependence of such quality metrics as PSNR and SSIM on training set size it was concluded that for the size of 25–30 scenes metrics are greater than 90% of the values claimed by the authors of the above-mentioned article for their model trained on 160 scenes. Also these metrics are compared for the model trained on the photos from the first camera, and model trained on new camera photos after their testing on a set of photos from the second camera. Despite the numerical results (PSNR/SSIM = 27.40/0.82 for the first and 30.06 / 0.83 for the second model) visual comparison of the quality of the details on the images for the model trained on a relatively large dataset from another camera are markedly better that confirms the cross-sensor generalization of the method of the mentioned article, as claimed by its authors.

Key words: image processing, denoising with neural network, computational photography

DOI: 10.31857/S0235009221010078

Cite: Popov N. I., Grigoryev A. S. Apparatnaya nezavisimost i tochnost neirosetevogo shumopodavleniya na izobrazheniyakh kak funktsiya obema obuchayushchikh dannykh [Hardware independence and accuracy of neural network denoising of images depending on training set size]. Sensornye sistemy [Sensory systems]. 2021. V. 35(1). P. 79–83 (in Russian). doi: 10.31857/S0235009221010078

References:

  • Aharon M., Elad M., Bruckstein A. KSVD: An algorithm for designing overcomplete dictionaries for sparse representation. Trans. Sig. Proc. 2006. V. 54 (11). P. 4311–4322.
  • Anaya J., Barbu A. Renoir - a dataset for real low-light noise image reduction. arXiv preprint arXiv:1409.8230. 2014. P. 6.
  • Bernd J. Digital Image Processing. 6th revised and expanded edition. ISBN 3-540-24035-7 Springer, Berlin, Heidelberg. New York. 2005.
  • Chen C., Chen Q., Xu J., Koltun V. Learning to see in the dark. CVPR. 2018. P. 3291–3300.
  • Dabov K., Foi A., Katkovnik V., Egiazarian K. Image denoising by sparse 3-d transformdomain collaborative filtering. TIP. 2007. V. 16 (8). P. 2080–2095.
  • Foi A., Trimeche M., Katkovnik V., Egiazarian K. Practical Poissonian-Gaussian Noise Modeling and Fitting for Single-Image Raw-Data. TIP. 2008. V. 17 (10). P. 1737–1754.
  • Gharbi M., Chaurasia G., Paris S., Durand F. Deep joint demosaicking and denoising. SIGGRAPH. 2016. V. 35 (6). P. 1–12.
  • Guo S., Yan Z., Zhang K., Zuo W., Zhang L. Toward convolutional blind denoising of real photographs. CVPR. 2019. P. 1712–1722.
  • Hasinoff S.W., Sharlet D., Geiss R., Adams A., Barron J.T., Kainz F., Chen J., Levoy M. Burst photography for high dynamic range and low-light imaging on mobile cameras. SIGGRAPH Asia. 2016. V. 35 (6). P. 1–12.
  • Kim D., Chung J.R., Jung S. Grdn: Grouped residual dense network for real image denoising and gan-based realworld noise modeling. CVPR. 2019.
  • Lehtinen J., Munkberg J., Hasselgren J., Laine S., Karras T., Aittala M., Aila T. Noise2Noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189. 2018.
  • Liu Z., Yuan L., Tang X., Uyttendaele M., Sun J. Fast burst images denoising. ACM Transactions on Graphics. 2014. V. 33 (6). P. 1–9.
  • Nam S., Hwang Y., Matsushita Y., Kim S.J. A holistic approach to cross-channel image noise modeling and its application to image denoising. CVPR. 2016. P. 1683–1691.
  • Perona P., Malik J. Scale-space and edge detection using anisotropic diffusion. TPAMI. 1990. V. 12 (7). P. 629–639.
  • Plotz T., Roth S. Benchmarking denoising algorithms with real photographs. CVPR. 2017. P. 1586–1595.
  • Ren W., Liu S., Ma L., Xu Q., Xu X., Cao X., Du J., Yang M.H. Low-Light Image Enhancement via a Deep Hybrid Network. TIP. 2019. V. 28 (9). P. 4364–4375.
  • Schwartz E., Giryes R., Bronstein A.M. DeepISP: Toward Learning an End-to-End Image Processing Pipeline. TIP. 2019. V. 28 (2). P. 912–923.
  • Simoncelli E.P., Adelson E.H. Noise removal via bayesian wavelet coring. ICIP. 1996. V. 1. P. 379–382.
  • Wang R., Zhang Q., Fu C., Shen X., Zheng W., Jia J. Underexposed Photo Enhancement Using Deep Illumination Estimation. CVPR. 2019. P. 6849–6857.
  • Zhang Y., Zhang J., Guo X. Kindling the Darkness: A Practical Low-light Image Enhancer. CVPR. 2019. P. 1632–1640.
  • Zhang K., Zuo W., Chen Y., Meng D., Zhang L. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. TIP. 2017. V. 26 (7). P. 3142–3155.