Denoising Images with Varying Noises Using Autoencoders

Abstract

Image processing techniques are readily used in the field of sciences and computer vision for the enhancement of images and extraction of useful information from them. A key step used in image processing involves the removal of different kinds of noises from the images. Noises can arise in an image during the process of storing, transmitting or acquiring the images. A model qualifies as a satisfactory de-noising model if it satisfies image preservation along with noise removal. There can be various kind of noises in an image such as Gaussian, salt and pepper, Speckle etc. A model which can denoise a different kind of noises is considered to be superior to others. In this paper, we have designed a model using autoencoder which can remove several kinds of noises from images. We have performed a comparative study between the accuracy of each kind using PSNR, SSIM and RMSE values. An increase in the PSNR and SSIM values was seen from the original and noisy image to the original and reconstructed image while a decrease was seen in the value of RMSE.