Publication: A Deep Learning Methodology for Automatic Assessment of Portrait Image Aesthetic Quality
Issued Date
2018-09-06
Resource Type
Other identifier(s)
2-s2.0-85057761362
Rights
Mahidol University
Rights Holder(s)
SCOPUS
Bibliographic Citation
Proceeding of 2018 15th International Joint Conference on Computer Science and Software Engineering, JCSSE 2018. (2018)
Suggested Citation
Poom Wettayakorn, Siripong Traivijitkhun, Ponpat Phetchai, Suppawong Tuarob A Deep Learning Methodology for Automatic Assessment of Portrait Image Aesthetic Quality. Proceeding of 2018 15th International Joint Conference on Computer Science and Software Engineering, JCSSE 2018. (2018). doi:10.1109/JCSSE.2018.8457381 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/45566
Research Projects
Organizational Units
Authors
Journal Issue
Thesis
Title
A Deep Learning Methodology for Automatic Assessment of Portrait Image Aesthetic Quality
Other Contributor(s)
Abstract
© 2018 IEEE. Generally, a traditional methodology to assess the aesthetics (appreciating beauty) of a photograph involves a number of professional photographers rating the photo based on given criteria and providing ensemble feedback minimize bias. Such a traditional photo assessment method, however, is not applicable to massive users, especially in real-time. To mitigate such an issue, recent studies have devoted on developing algorithms to automatically provide feedback to photo takers. Most of such algorithms train variants of neural networks using ground-truth photos assessed by professional photographers. Regardless, most existing photo assessment algorithms provide the aesthetic score as a single number. From our observation, users typically use multiple criteria to justify the beautifulness of a photo, and hence a single rating score may not be informative. In this paper, we propose a novel Fine-tuned Inception with Fully Connected and Regression Layers model which gives five attribute scores: vivid colour, colour harmony, lighting, balance of elements, and depth of field. T his s olution i ncorporates t he p re-trained inception model which is the state-of-the-art model for processing images. Our proposed algorithm enhances the existing state-of-the-art by fine-tuning the parameters, introducing fully connected layers, and attaching the regression layers to compute the numeric score for each focus attribute. The experimental results show that our model helps to decrease the mean absolute error (MAE) to 0.211, benchmarking on the aesthetics and attributes datasets provided in the previous studies.