A Deep Learning Approach for Gaussian Noise-Level Quantification

Abstract

Image noise and its reduction have been of active interest over the last few years due to the increasing application of Image processing in various domains such as medical, video processing, robotics, etc. This paper presents a Deep Learning (Convolutional Neural Network) approach to classify an image into different levels of Gaussian noise it has. Most of the image-denoising models are based on the assumption that the images to be processed have some level of noise in them. The proposed model evaluates whether the image has noise present in it and also quantifies the amount of noise the image has. The MSRA 10K dataset has been used to train, test, and validate the model. Ten levels of Gaussian noise have been introduced to the dataset which represent the ten classes for the model and one class with a clean image. The model has achieved an accuracy of 96%. The proposed model has proven to be very effective in identifying the level of Gaussian Noise present in the image.