https://rit.zoom.us/j/2916170382
June 8, 2020 at 1:00pm
Michal Kucer
Ph.D. Thesis Defense

Chester F Carlson Center for Imaging Science

Ph.D. Thesis Defense

Michal Kucer

Representations and representation learning for image aesthetics prediction and image enhancement

Advisors: Dr. David Messinger

Monday, June 8th, 2020, 1:00pm

Location: https://rit.zoom.us/j/2916170382 

 

Abstract: 

With the continual improvement in cell phone cameras and improvements in the connectivity of mobile devices, we have seen an exponential increase in the number of images that are captured, stored, and shared on social media. For example, as of July 1st, 2017, Instagram had over 715 million registered users which had posted just shy of 35 billion images. This represented approximately seven and nine-fold increase in the number of users and photos present on Instagram since 2012. Whether the images are stored on personal computers or reside on social networks (e.g. Instagram, Flickr), the sheer number of images calls for methods to determine various image properties, such as object presence or appeal, for the purpose of automatic image management and curation. One of the central problems in consumer photography centers around determining the aesthetic appeal of an image and motivates us to explore questions related to understanding aesthetic preferences, image enhancement and the possibility of using such models on devices with constrained resources.

In this dissertation, we present our work on exploring representations and representation learning approaches for aesthetic inference, composition ranking and its application to image enhancement. Firstly, we discuss early representations that mainly consisted of expert features, and their possibility to enhance Convolutional Neural Networks (CNN). Secondly, we discuss the ability of CNNs, and the different architecture choices (inputs size and layer depth) in solving various aesthetic inference tasks: binary classification, regression, and image cropping. We show that if trained for solving fine-grained aesthetics inference, such models can rival the cropping performance of other aesthetics-based croppers, however they fall short in comparison to models trained for composition ranking. Lastly, we discuss our work on exploring and identifying the design choices in training composition ranking functions, with the goal of using them for image composition enhancement.