Mobile Deployment of NoiseNet: Noise Characteristics Assessment in Real-World Images

Processing of real-world images permanently faces noisy conditions during data acquisition. Distortions in images introduced by noise can decrease efficiency of subsequent treatment significantly. To suppress the noise effectively or to get the rationale of denoising expedience, noise characteristics should be assessed in an accurate way. If not, inaccurate estimation of noise properties can affect denoising performance like oversmoothing or insufficient noise attenuation. This paper considers the use of deep convolutional neural network NoiseNet and other tools to estimate characteristics of signal-dependent noise in real-world images acquired by mobile devices. Noise characteristics' estimation is conducted in different conditions of photo capturing and using mobile devices from different manufacturers. It is shown that, in most cases, the noise in images is spatially correlated. Additionally, it is noticed that image content greatly influences noise variance estimation accuracy, especially in cases of heterogeneous or highly-textured images. Ready-to-use trained model of NoiseNet and the demo Android application with full dataset are available at https://github.com/radiuss/NoiseNet.