How computers are learning to read and translate ornate handwritten Tamil

Solving a curved script puzzle

Tamil is one of the oldest living languages in the world. It uses flowing, round letters with lots of curved loops. When people write it by hand, they often add personal flourishes. These extra swirls easily confuse computers. Standard vision systems often misread the symbols completely, which makes accurate Tamil handwriting translation very difficult.

A modelling study has tested a new system called THCRT to solve this problem. Instead of feeding messy scans straight into a neural network, the system cleans up the image first. Raw handwriting passes through filters to boost local contrast and remove speckled noise. Many writers let their text drift up or down. To fix this, a tool called CPFCA uses polar coordinates to unwrap curved lines into straight rows in just 12.4 milliseconds.

Smoothing out decorative lines

Next comes the trickiest part: uneven pen strokes. Writers often press down hard or add sweeping loops that mask the true letter shape. The system applies a fuzzy logic filter called BEMFFIS. This filter evens out stroke thickness across the writing before thinning each shape down to its bare skeleton.

Think of this filter like ironing a wrinkled shirt before measuring it. Smoothing the cloth reveals the true seams underneath. The pipeline then sends the thinned letters into a convolutional network called GAE-DCSSNN to identify each character. On the 16,184-sample TOHWR test set, the system achieved 98.23% accuracy, correctly identifying 15,898 words. Its character error rate was just 1.12%. On a second set of 4,940 samples, it reached 97.61% accuracy. Taking away the fuzzy stroke filter made the character error rate jump from 1.330% to 1.980%, proving that stroke cleanup is vital.

Turning script into speech

Reading a word is only the first step. Tamil builds rich meanings by gluing prefixes and endings onto root words. A machine that cannot spot these parts will fail to translate the text. To fix this, the pipeline uses mBBDART, a slimmed-down translation model guided by dependency parsing. Dependency parsing maps out the grammar rules that link roots to their word endings.

On a 1,200-sample expert reference set, this approach achieved a BLEU score of 0.438 and a METEOR score of 0.512. Adding grammar rules gave the model a 40.4% relative boost over an unguided base model. The team also found that human errors in standard benchmarks held machines back. Correcting mislabelled test data boosted scores across six neural setups by an average of 2.14 percentage points.

What we still do not know

The BLEU score of 0.438 shows how hard it is to translate single words without full sentence clues. Earlier studies also used private test sets, making direct comparisons suggestive rather than final. The translation tool has only faced isolated words so far, not full paragraphs. We do not yet know if it can handle long texts on large benchmarks like FLORES or Samanantar. Researchers also need to test whether the line-straightening tools work on related scripts like Telugu or Kannada without manual tuning.