Convolutional Autoencoder for Image Denoising: A Compositional Subspace Representation Perspective

Abstract

This study explores a convolutional autoencoder for image denoising with a proposed compositional subspace method. This modeling approach presents a structural and rigorous mathematical abstraction to understand a convolutional autoencoder's functional computation layers. The theoretical basis is that the best way to model a complex learning function is by using a composition of simple functions to form a multilayer successive cascaded function for complex representation. The proposed method has experimented with the Fashion-MNIST dataset. Experimental results are discussed and were consistent with the theoretical expectation.