Denoising face images using convolutional autoencoders

Abstract

In this work, we articially introduce three types of noise (additive gaussian noise, multiplicative noise and salt and-pepper noise) to black-and-white images of human faces and then try to remove the noise by using a specic type of neural network architecture known as an autoencoder. To boost the e ciency of this network, we use convolutions, rectied linear units and change the cost function to the Structural Similarity Index Measure (SSIM), an index that measures similarity between images, which we then compare to the more usual mean squared error (MSE) cost function.