The project "Architectural Styles Recognition" aims to differentiate pictures between three different architectural styles. We chose the categories: Neo-futuristic, Gothic, and Brutalist. Due to their clear differences in building styles and materials the Artificial Intelligence algorithm could more reliably classify the images into the applicable style category.
First tests showed that, even with a limited amount of training data (~16 images per category), the program was mostly able to distinguish between the chosen styles. The final model, with over a hundred training images per category, further increased the accuracy. The training data included pictures showing whole buildings from the outside as well as close-ups of different facades. We also occasionally included interiors. We collected the images mainly through different search engines, ensuring a geographically diverse sample, and added photographies from personal archives.
We decided against the inclusion of a fourth category (i.e. 'None of the above'), as we were curious on how the program would classify unexpected data in the peculiar range of Gothic, Brutalist, and Neo-futurist architecture.
While the final results that can be obtained are largely correct, the model can be challenged by images that do not show stereotypical buildings or images corresponding to the three architectural styles. For example, due to cleaner lines and less detailed characteristics, the Brutalist and the Neo-futurist styles can sometimes be confused by our model, depending on the input picture's angle and the particular style of the building being depicted.
Interesting results are obtained when the input image does not belong to one of the three chosen styles but rather to other styles, like Islamic architecture, where buildings are often associated either to the Gothic or the Neo-futurist style, or, contrary to what its name might suggest, the Futurist architectural style is often recognized as Brutalist.
To test our model drag an image from your computer into the black square to see the result written on the bottom center of the image (it might take a few seconds for the image to load):
The model can also be tested using the Teachable Machine interface by uploading files or by using the webcam. At the end of the article you can find some images which can be used for testing.
You can download the following images to test the model. These images were not included in the training dataset.