Improving the State of the Art
Updated: 3 days ago
The client required an extremely versatile platform for searching a large image database; from a concept to a working tool which was powerful enough to be presented for a funding round, and one which would form the basis of the future company. This not only involved the project design, architecture, database, development, deployment and hosting, but the curation of the data and images for the proof of concept as well as shaping the requirements.
The proposal was to provide a comprehensive tool combining three primary approaches, focussing primarily on renaissance era artworks;
image & painting analysis techniques, implemented by datagram,
traditional physical artwork examination data, such as X-ray spectrometry, pigment information collated by the client
And art history, for artists and paintings
The product layers that we implemented included the backend graph database, the data injestion pipelines, the web API, the frontend data visualisation and the user interface, including authorisation, deployment and reliability testing.
We did significant research, reviewing multiple papers in the domain. The techniques that we investigated included image search routines, various entity extraction techniques, CNNs, fast retrieval and vector search, knowledge graph ontologies among many others.
At the same time as ensuring that the techniques we used were sufficiently powerful, we had to also ensure that the output was suitable for the customers' requirements, and fast enough to give responsive results. We realised there was no 'one-size' of painting, and some techniques perform much better with some styles than others.
The solution was refined through weekly meetings with the client over a 3-4 month period, resulting in the 'finished' product.
Technologies; hosted on Microsoft Azure, with data tables for quick access data. Standalone Graph database, Cypher query language, NuxtJS for the frontend, and Python for the webserver.