Reconstructing Noised Images of Fashion-MNIST Dataset Using Autoencoders

In this paper, we delve into the potential of a structured latent space for enhanced data representation, particularly advantageous for tasks such as categorization with sparsely labeled data, effective label selection, and class morphing. Our research encompasses evaluations conducted on diverse benchmark datasets, including ImageNet, MNIST, Fashion-MNIST, DeepFashion2, and a collection of 3D human forms, demonstrating the versatility of our approach. Within the context of Denoising Autoencoder methodologies, our primary aim was to grasp and acquire input representations capable of disregarding inconsequential changes and comprehending the entire image. We introduced noise to the images and employed denoising autoencoders to reconstruct the original images, eliminating noise. Our findings show that a learning rate of 0.1 and 1000 hidden layer nodes were optimal for our model. The model excelled at reconstructing images from the test set when trained with 20% noisy images but struggled to generalize with 70% noise due to overfitting. Training with noisy images yielded improved testing accuracy, and increasing the number of samples for the classification layer was beneficial. However, our model's performance fell short in comparison to Support Vector Machines (SVM).