The Cretogram Dataset for Isolated Character Recognition in Archaic and Classical Crete
This paper introduces Cretogram, a new dataset of 11,611 manually annotated isolated epigraphic characters from Archaic and Classical Crete, and evaluates its utility for computational recognition using HOG-SVM and deep learning models, achieving a top macro-F1 score of 95.53% with ResNet-18 while highlighting specific character confusions inherent to the script's local variations.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The study of ancient writing often feels like trying to read a book where the pages have been torn out, the ink has faded, and the letters have been reshaped by centuries of weathering. In the field of epigraphy, which is the study of inscriptions carved into stone or cast in metal, scholars have long relied on their own eyes to decipher these marks. They compare the curves of a letter carved in one Greek city to the same letter carved in another, looking for subtle differences that reveal where a stone was made or when it was written. For the island of Crete, this task is particularly complex. During the Archaic and Classical periods, the island was not a single, unified writing culture but a collection of local traditions. Each city-state developed its own slight variations in how it drew its alphabet. A letter that looks like a standard "A" in one town might have a different shape in a neighboring valley. These small variations are the key to understanding the history and geography of the ancient world, but they are also incredibly difficult for a computer to learn, because the computer must distinguish between a genuine difference in style and a simple smudge or a broken piece of stone.
A team of researchers at the Aristotle University of Thessaloniki has tackled this challenge by creating a new digital resource called Cretogram. Their goal was to teach a computer to recognize these ancient Cretan letters in isolation, without the help of the surrounding words that usually give a human reader context. To do this, they did not simply scan old photographs; they went back to the scholarly drawings and transcriptions that experts have spent decades perfecting. From 187 specific inscriptions, they carefully extracted 11,611 individual characters. Each character was cleaned up, standardized to a uniform size, and labeled with its correct identity, but crucially, the researchers kept a digital link to the original stone it came from. This connection is vital because it allows the computer to be tested in a realistic way: the researchers trained the system on some inscriptions and then asked it to identify letters from completely different inscriptions it had never seen before. This method ensures that the computer is learning the actual shapes of the letters, not just memorizing the specific style of a single stone cutter.
The results of this experiment show that modern artificial intelligence can indeed master these ancient forms, though it faces specific hurdles. The researchers tested three different approaches to the problem. One method used a traditional technique that looks at the edges and angles of the shapes. A second method used a compact, custom-built neural network, which is a type of computer program designed to mimic the way the brain processes visual patterns. The third method used a much larger, pre-trained system that had already learned to recognize millions of everyday images, such as cats and cars, before being adapted to these ancient letters. The largest system performed the best, correctly identifying the letter in nearly 98 percent of the test cases. It successfully distinguished between the many local variations of the Cretan alphabet, proving that a machine can learn to see the subtle differences that define these ancient writing traditions.
However, the study also revealed where the machine still struggles, highlighting the limits of looking at a single letter in a vacuum. The computer frequently confused certain pairs of letters, such as epsilon and digamma, or gamma and lambda. In many of these cases, the confusion was not a failure of the software but a reflection of the ambiguity in the source material itself. Sometimes a letter is carved with a faint or broken stroke, or its shape depends entirely on the letters next to it to make sense. Without the context of the full word, even a human expert might hesitate. The researchers found that when they allowed the computer to see letters from the same inscription during both its training and testing phases, its performance improved slightly. This suggests that the computer benefits from seeing the consistent style of a single stone cutter, but the strict test of recognizing a letter from a completely new stone remains the most difficult and most important challenge.
The significance of this work lies in the dataset itself. By making these 11,611 images available to other researchers, along with the code used to test them, the authors have provided a common ground for future studies. They have shown that while computers can now read these ancient letters with high accuracy, the most difficult cases often require the kind of historical context that only a human scholar can provide. The project does not claim to have solved the mystery of ancient Cretan writing, but it has built a powerful new tool for examining it. The researchers have demonstrated that by combining careful historical scholarship with modern machine learning, it is possible to preserve and analyze the unique visual fingerprints of local alphabets that have survived for over two thousand years. This approach allows the digital world to respect the complexity of the past, acknowledging that some answers lie not just in the shape of a single letter, but in the story of the stone it was carved upon.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.