Supervisors: Dr. Martin Danelljan, Dr. Radu Timofte, Prof. Luc Van Gool
Over the past decade, deep neural networks have taken over numerous computer vision tasks. However, their robustness is still heavily depedent on the availability of large datasets. Furthermore, these datasets must also be qualitatively annotated. Among the various tasks of computer vision, image quality assessment (IQA) has become increasingly important. Indeed, being able to quantify quality of images is of high importance for various problems such as image enhancement, denoising, super resolution, etc. While quantifying similarity is straightforward for various types of data (vectors, text, etc.), this same task reveals itself more difficult for visual patterns - a task at which humans excel. In this work, we focus on perceptual image quality assessment (IQA), which attempts to predict human perception of distorted images, and exclusively focused on the full-reference IQA problem. Getting human-annotated datasets is very expensive and often limits the capabilities of trained models. In order to tackle this problem, we present a method to synthetically generate training data of different quality levels and we test our method on two existing datasets. We also explore and analyze the newly introduced Berkeley-Adobe Perceptual Patch Similarity (BAPPS) dataset and Learned Perceptual Image Patch Similarity (LPIPS) metric. Experiments show that our approach improves the generalization performance of the learned IQA net- works, while preserving the accuracy. By increasing the variety of training samples, we tend to limit over- fitting, which is an omnipresent challenge when working with neural networks. We provide through our method a way to exploit any existing dataset almost effortlessly to tackle this challenge.