Comparative Study of Autoencoders Used for Noisy Images Restoration

The paper is devoted to solving the problem of restoring noised images by convolutional autoencoders. It is provided a comparative research of three types of convolutional autoencoder architectures: the standard architecture based on convolution and deconvolution operations; the architecture based on Residual-blocks; the architecture based on Residual-blocks using a discriminator network. The effectiveness of the autoencoder architectures have been evaluated on synthetically noisy color images. Subjective and objective metrics as MSE, RMSE, PSNR, SSIM were used to evaluate the quality of the restored images. In the course of a comparative study, some recommendations were drawn on which of the explored autoencoders to choose providing the best effectiveness restoring the quality of distorted images.