Automatic Theme Detection from Social Media Images using Deep Learning (DEEPTHEME)

DigitalMR proposes to investigate methods of determining themes in collections of images
that accompany social media posts. The methodology is inspired by recent advances in deep
learning that have benefited from the availability of large training data sets along with
increased computation power through heterogeneous computing. The concept operates
unsupervised and leverages the “deep learning framework” to determine themes and establish
their relevance to brands or organisations using hierarchical structures. The concept also
assigns labels to identified themes (topics) and determines ways that describe the theme such
that it can be applied to market research and insight management tasks, such as sentiment and
semantic analysis. The target outcome of the project is to discover the potential and to reserve
the capability of theme detection in image collection for commercial applications. This
capability will ultimately enhance listening247, a social media listening and analytics system
(text based) whose effectiveness has been proven by a range of private and public sector
organisations. The core R&D tasks include methods for learning imbalanced datasets, deep
architecture selection, deep learning via dedicated classifiers and ensemble formulations. The
project will make use standard datasets for training and testing. The DigitalMR team will
benefit from systematic input by a specialist UK academic team (subcontractor) as well as
user feedback from corporate challenge partners.