Artificial intelligence clears up questions about porcelain artefacts

Imagine that you manage a large collection of valuable historical porcelain artefacts that generate a lot of interest from the public. You want it to be accessible, but dealing with all the questions about the objects is a labouring task. That's where the seed of a new, user-friendly service with a modern AI solution started to grow. Plus a fortunate circumstance, when two partners came together - RISE and the Rörstand Museum.

The challenge of open collections

Rörstand Museum is located a stone's throw from Lidköping's main square, in the old porcelain factory that was a major employer of more than 1,700 employees during its golden days. It houses a gigantic collection spanning 300 years of porcelain history, with more than 1,800 items in storage. Much of the collection is open to the public six days a week.

- "We started thinking about how we could make the collections available to those interested in porcelain in an even better way, and that's when the idea of AI came up," says My Johansson Dufva, CEO of Rörstrand Museum.

With open collections come questions - many questions! With that challenge plus a curiosity to develop, the museum started an AI project just over three years ago together with researchers at the RISE Centre for Applied AI.

- "We will develope AI technology to create a useful service with a prototype available this spring," says Olof Mogren, research manager at the Centre for Applied AI at RISE.

The plan is to use AI to respond to the large number of questions, which require considerable resources to answer manually via e-mail. Many questions require staff to look for the answers in the physical warehouse - a detective work that requires time, commitment and high competence. Often the response process is slow and answers lead to new follow-up questions.

- "We receive about 1500 questions per year and answering all of them requires one employee. We saw both an opportunity to streamline the work and make the response management faster, but also give more interested access to the answers," explains My.

The solution

Common and recurring questions to the museum about artefacts range from who the artist is, the name of the porcelain design, the production year, to which parts are in the same collection.

The project began by working with RISE and researchers at the Centre for Applied AI to set up an AI model that was trained on professional photographs of more than 5,000 objects in a digital object management system.

The importance of good training data 

An AI model based on deep learning was trained on the images in the digital collection. From Rörstand's customer perspective, it was most important for the model to learn to recognise the artist and year of manufacture.

- "We model the age of the objects and who created them. The model can also determine what type of object it is," Olof explains.

Using image analysis, the research team at RISE therefore began to examine the underside of the porcelain objects. This often contains useful information in the form of stamps and logos. An object recognition model (YOLO) recognises the objects and marks boxes where it finds things. Once the logo is recognised, it is easier to determine the age of the object, as Rörstrand's logo has evolved over the approximately 300 years of production.

- "By recognising different logos, it is possible to date the item by decade," says Olof Mogren.

Challenges and lessons learned

What was discovered in the first step was that the data from the museum's collections was not sufficient to make reliable analyses. However, My and her staff already had an unexpected positive side effect - new lessons learnt! The AI made its own conclusions and gave the staff new insights, such as the fact that most of the digitised material was represented by only eight artists.

- "From a researcher's perspective, however, the same statistics can highlight a difficulty, namely that the training data represents too few artists," Mogren adds.

Another problem was that tableware has not always been named historically, so there are no answers for either staff or AI. The third challenge was that the training images were too perfect (they were created with a professional camera, with high sharpness, and against a solid-coloured background). Paradoxically, this may mean that the training data is not sufficiently representative, and a resulting AI solution will not be robust enough to be used on images that porcelain enthusiasts have taken in their homes.

- "Training an AI model on images requires a large number of images. "We are building a robust model that will be able to interpret different image quality, so we need to train it on images taken in different lighting conditions, with different noisy backgrounds and even blurred images," Olof explains.

Thus, even if a data set is sufficient, it must be representative and provide examples of the images that the end user has when asking questions about the objects.

Areas of improvement and next steps

The next step was therefore to increase the representative data set. Therefore, users were encouraged to upload their own images of porcelain artefacts to complement the professional studio images already in the system. This was done by building a platform to engage the public, who were given the opportunity to submit their own images and enter the nature of the object. The goal was to make the AI model more robust.

- "There is always a degree of uncertainty in the answers from an AI model. In this project, we use uncertainty quantification models, where the AI system can indicate how confident it is. To make it clear to the end user, the AI tool will respond with a language that is adapted to the uncertainty quantification; for example, if it is unsure of an analysis, it can indicate which two designers it is weighing between or give an approximate time span for the year of manufacture," explains Olof.

The project contributes to increased competitiveness and open society

Through this type of collaboration between researchers at the Centre for Applied AI and stakeholders with real societal challenges, the development of services in Sweden is strengthened, contributing to a competitive and open society. The project will result in a fully usable solution that Rörstrand Museum will be able to use in its work from this spring. The project not only strengthens the museum's role as a knowledge bank, but also benefits from the great commitment and expertise of porcelain collectors.

- "It has been extremely exciting to collaborate with researchers at RISE, and if we succeed in getting the AI to answer questions that we cannot answer today, it will be extremely valuable. It will also be a solution that can be applied to other collections, and it's great to be at the forefront of that development," concludes My.

Read more: Science of AI