Application of deep learning models for image denoising

Abstract

This paper addresses the issue of image denoising using neural networks. Experimental results of training and testing neural network models on images with artificially added noise are presented. In today's digital world, image quality plays an important role in various fields, including medicine, science, entertainment, and communication. Noise caused by imperfect equipment, problems in transmission channels, or random physical factors significantly reduces the efficiency of image processing and perception. To train and test neural network models, a suitable set of images is required. In this study, we created training, validation, and test sets containing original images and images with added noise. Three sets of images from the www.kaggle.com were used to form them. The first set contained butterfly images, the second set comprised athlete images, and the last set included human face images. The training set was formed as follows. 1000 images were randomly selected from each of the three sets and scaled to a size of 128x128. From these, 30 images were randomly selected and subjected to 10 levels of noise addition, resulting in 300 images for each of the 6 noise levels. In total, the training set contained 1800 images. The validation sample was formed in a similar way, but was half the size. The test set consisted of 100 original images with three noise levels added. The following types of noise were used: Gaussian, Poisson, salt and pepper, Laplace, mixed, and quantization. Three models were trained for image denoising: two convolutional auto-encoder models (with 3 and 6 convolutional layers in the encoder and decoder, respectively) and a modified U-Net model in which the output layer and loss function were adjusted. The performance of the trained models was compared with each other and with classical filters on test images. The experimental results show that the U-Net outperforms autoencoders in noise removal from images. In comparison to classical filters, the U-Net model reveals competitive outcomes. It shows better quality in the case of Gaussian, Laplace, and quantization noise, especially in the context of preserving structural information (higher SSIM).