Predicting aesthetic value of photographs is a challenging task for any computer system. Even humans experience a lot of difficulties when explaining why a given picture is perceived as aesthetically pleasing. This is why it does not seem to be possible to solve this challenge by defining a set of simple rules. Instead, we believe that this problem can be addressed by referring to a crowd-sourced dataset of photographs with corresponding popularity score which we treat as aesthetic metric proxy. Training a machine learning algorithm using this dataset appears to be a more practical approach to the problem and we follow this methodology here. More precisely, in this paper, we assess the aesthetic value of a photograph using only its pixels. Building up on the successful applications of deep convolutional neural networks in other related domains, such as image recognition [7, 8, 10], we propose such networks to address also this problem. Our method is strongly inspired by previous research on similar problems, which concluded that deep convolutional neural networks perform well in such cases [2, 3, 5, 6]. One could imagine a wide variety of applications for a system solving the problem stated in the paper. First of all, such a system can significantly improve the workflow of every photographer. By preselecting or suggesting the best photos from a defined set we can save a lot of time, storage space and network traffic. Usefulness of the system could be furthermore improved by combining it with some way of detecting similar photos. That results in easy removal of duplicates, which in turn saves photographer the time that he would spend on selecting the best frame out of a given set. To provide another example, one could even imagine a camera or a post processing software using such system to suggest and perform automatic image enhancements, like e.g. exposure compensation or even cropping. Our paper provides the following cont[...]
Wyniki 1-1 spośród 1 dla zapytania: authorDesc:"Maciej Suchecki"