Unsupervised Learning for 2D Image Texture Enhancement

One disadvantage of supervised learning is the capability of the neural network to only recognize objects labeled during the training phase. In order to identify objects outside of the training dataset, unsupervised learning is needed. In this paper, autoencoder is used as the unsupervised neural network and it also functions as a 2D image texture enhancer. Autoencoder is capable of clustering and dimensional reduction, where reducing dimensions of images can be categorized as feature selection and feature extraction. The role of autoencoder as an unsupervised network is to train different High-Resolution (HR) images and preserve texture features that can be used later to enhance Low-Resolution (LR) input images and process them into an enhanced 2D image (E2D). Low-resolution images due to degradation or blurring can be enhanced with the proposed algorithm under 2 parts of autoencoder: encoder and decoder. The encoder stage downsizes the HR dataset while maintaining the extracted features during the training phase. The decoder stage makes use of the extracted feature and denoises the LR input image to produce an E2D image. 100 LR images were tested with a 100^% success rate of E2D enhancement.

I. Introduction

In recent years deep learning has been the core of artificial intelligence (AI), especially in computer vision fields. In computer vision technology, convolutional neural network (CNN) is the popular choice in making the model for object detection because of its performance in training many images. However, this model is mostly used in supervised learning, in which images are trained to identify a set of objects into classes. The disadvantage of supervised learning is the cost and time involved in selecting the training data. This includes pre-defining and labelling the data into different classes, which limits the categories and class selection. In comparison, unsupervised learning does not require the training data to be labelled, which reduces the cost and time for classification [1].