Performance Analogy of Autoencoders for Image Denoising

Image Denoising is one of the current day challenges in the field of image processing. Noise in the image is the existence of artifacts that do not start from the original picture content. Autoencoders have made considerable progress in image denoising because of their ability to reconstruct image inputs. Autoencoders for denoising images usually accepts a degraded image as input and is trained to predict the reconstructed image as its output. An effort has been made to analyse the functioning of selected autoencoders for image denoising, such as Basic Autoencoder, Denoising Autoencoder and Convolutional Denoising Autoencoder using the Modified National Institute of Standards and Technology database (MNIST) dataset. The performance of these models is evaluated and compared using Structural Similarity Index Measure, Peak Signal to Noise Ratio evaluation metrics. The strengths and weaknesses of each autoencoder are discussed and the factors that influence their performance are explored. Qualitative and quantitative analysis of the dataset provides an insight of effectiveness and efficiency of each autoencoder in denoising images. Observations show that Convolutional Denoising Autoencoder out-performs other autoencoders in terms of the evaluation metrics and better reconstructed image. This study can act as a base for researchers and practitioners to identify suitable autoencoders for denoising images in real-life applications.