Assessing the Aesthetic Quality of an Image With Deep Learning; An Exploratory Research for the United Kingdom
By Zardalidis Apostolos
This master thesis focusses on the correlations of aesthetic value and the high-level properties of an image. High-level properties of an image are directly associated with its contained objects. Those objects can be categorized in order to conduct statistical analysis and understand the perceived beauty of each class. Two major datasets are needed to proceed with this research. One contains the aesthetic value of geo-tagged images given from a crowd-sourced project and the other contains the landscape classes provided from a supervised classification. The secondly mentioned dataset was used to train a deep neural network (DNN) model for object recognition and identification of patterns of the existing landscape classes. The other dataset was forward propagated in the DNN model to collect predicted class-scores for each image. These class-scores were derived from the model’s attempt to match the learned patterns of objects with those contained in the forwarded dataset. The aesthetic value of each landscape class was correlated with the geographic location and urban – rural areas. These results were visualised through map representations.