Adaptive image denoising using a deep neural network with a noise correction map

Smartphone camera is becoming the primary choice for photography among general users due to its convenience and rapidly improving image quality. However, it is more prone to noise compared to a professional DSLR camera due to a smaller sensor. Image noise, especially in low-light situations, is a critical problem that must be addressed to obtain high quality photos. Image denoising has thus remained an important low level vision topic over years with both traditional and learning based techniques used for mitigating this problem. We propose an adaptive Deep Neural Network based Noise Reduction (DNN-NR) algorithm to address the denoising problem in smartphone images. Image noise was modeled from photos captured under different light settings using a Poisson-Gaussian noise model which better approximates the signal-dependence (photon sensing) and stationary disturbances in the sensor data. Using this noise model, synthetic noisy datasets were prepared to mimic photos captured under varying light conditions and train the network. A noise correction map based on camera and image information like ISO, vignetting map and image gray level was provided as an input to the network. This correction map provides an indication of the local noise level to help the network adaptively denoise photos. Experimental results show that our adaptive neural network based denoising approach produced a significantly better denoised image with higher PSNR and MOS quality scores in comparison to a standard denoising method like CBM3D across varying light conditions. In addition, using a locally varying noise map helped in preserving more detail in denoised images.