Architectural Heritage Recognition with Deep Learning: An ANN and CNN Approach

History is often written in stone, but for centuries, understanding the stories carved into that stone has relied on the slow, careful work of human eyes. When a historian or architect looks at a building from the distant past, they must identify its style—whether it is the soaring arches of the Gothic era or the intricate columns of ancient Persia—to understand its place in time. This manual process is not only time-consuming but also difficult to scale when faced with thousands of images scattered across archives. The visual complexity of these structures, with their unique textures, shapes, and materials, makes it hard to organize them systematically. In recent years, a branch of computer science known as deep learning has offered a new way to tackle this problem. These systems are designed to learn from data much like a human brain learns from experience, allowing them to recognize patterns in images that might be too subtle or numerous for a person to sort through efficiently. By teaching computers to see the specific details that define a building's era, researchers hope to create tools that can automatically catalog and preserve our shared architectural heritage.

In a recent study, researchers Majid Ali Baig and Sarah Abbas from the University of Management and Technology set out to build such a tool. They focused on the challenge of automatically sorting images of historical buildings into their correct architectural styles. To do this, they developed a computer program based on a type of artificial intelligence called a convolutional neural network. You can think of this network as a digital system that looks at an image and breaks it down into layers of features, starting with simple lines and shapes and gradually building up to complex patterns that define a specific style. The team started with a large collection of historical building photos available online, which contained images representing twenty-five different architectural styles. For their initial experiment, they selected four distinct styles to test their system: the ancient Achaemenid style, the medieval Gothic style, the modern Deconstructivist style, and the ornate Art Nouveau style.

The researchers first prepared the images for the computer to read. They converted every photo into a grayscale image, removing color to focus purely on the shapes and structures, and resized them all to a uniform size of 256 by 256 pixels. This standardization ensured that the computer was comparing apples to apples, rather than getting confused by different image sizes or color variations. They then fed these processed images into their custom-built neural network. The system was designed with several layers that worked together to extract the most important details from each picture. It used mathematical techniques to adjust the importance of different features, ensuring the system learned the right patterns without getting confused by noise or irrelevant details. To make the system as effective as possible, the researchers used an automated tool to test thousands of different settings, searching for the perfect combination of parameters that would allow the computer to learn the fastest and most accurately.

The results showed that the computer could indeed learn to distinguish between these complex architectural styles. After training on the data, the system achieved a high level of accuracy, correctly identifying the style of a building in the vast majority of cases during its training phase. When tested on new images it had never seen before, the system maintained a strong performance, correctly classifying the architectural style in about 85 percent of the cases. The researchers also measured how well the system performed using specific metrics that look at both how often it was right and how often it missed a correct identification. The system showed a balanced ability to find the right style without making too many mistakes, with a score that reflected a solid understanding of the visual differences between the styles. However, the study also revealed that the system was not perfect; it occasionally confused the Gothic style with the Romanesque style, suggesting that some architectural features are so similar that even advanced computer models can struggle to tell them apart.

This work demonstrates that deep learning can serve as a powerful assistant for historians and preservationists. By automating the initial sorting of thousands of building images, the system can help organize vast collections of architectural records much faster than a human could. The researchers noted that their approach is scalable, meaning it could be expanded to handle the full range of twenty-five styles in their dataset, not just the four they tested. While the current model is a significant step forward, the authors acknowledge that there is still work to be done. Future efforts will likely focus on refining the system to handle more styles and improving its ability to distinguish between very similar architectural periods. For now, the study proves that computers can be taught to recognize the visual language of history, offering a new way to document and understand the built heritage that surrounds us.