Spotlight on scientific image analysis

Scivision, Connecting computer vision model developers to image data providers

Scivision is an open-source computer vision toolkit. Scivision provides researchers with an interface to search, discover, and integrate datasets and AI algorithms from a wide range of research areas. Scivision grew organically out of the Environment and Sustainability theme to support our objectives for open and reproducible research by creating data pipelines that limit duplication of development work. Scivision is collaboratively co-designed by researchers from within the Turing as well as external partners. Embracing the interdisciplinary model of data science, Scivision brought together researchers from different domains who use computer vision to study image data including plant phenotyping experiments, biomolecular characterisation by electron microscopy and satellite-imaging-based landscape monitoring. Their specific focus in Scivision was to address the common need to identify generalisable digital infrastructure that could make it easy for them to apply AI/ML methods across various data challenges. Their collaboration led to the development of an analytic framework that is agnostic to the choice of the algorithm as well as the intended application of the outputs.

The overarching aim of this project is to develop and deploy technological solutions for computer vision-based image analysis problems commonly found across the sciences and humanities, with a specific focus on the environment. Experts in diverse scientific domains use Scivision’s general-purpose Python toolkit to improve their framework for collecting and analysing data. The Scivision platform contains searchable catalogues of pre-trained computer vision algorithms and image datasets, allowing users to test out a range of algorithms in order to choose the one that best fits their data and application. The core features of Scivision are 1) the Scivision catalogue, containing pre-trained computer vision models and datasets from science and the humanities, and 2) the Scivision model and data source API, a simple, standard interface to models and data that works with the entries in the catalogue.

Scivision members have developed tools and infrastructure that democratises AI/ML and encourages exploratory use of existing toolkits for their applications from one field to another. The Scivision team collaborates with researchers in the UK to build examples showing how the tool can be used to perform tasks such as coastal vegetation detection and the identification of trees from satellite images. Researchers with any level of programming skills can pick up these examples and try them out with minimal intuitive changes to the executable code.

Testing and demonstration of Scivision’s features have been made possible via interactive Jupyter notebooks from research projects at the Scivision Gallery along with tutorials for new users. These notebooks can be run in the cloud through a service called Binder, which is supported in part by cloud infrastructure provided by the Turing. A standard graphical interface providing images and visual cues further allows environmental researchers, along with other non-coding user communities, to search for datasets and models compatible with their data of interest. Focusing on tackling real-world challenges, Scivision provides reproducible and reusable solutions for computer vision problems commonly found across the sciences and humanities. By enabling exploratory analysis, Scivision will help guide decisions on AI tools and underlying algorithms to be applied in the wider contexts of environmental research.