Fuzzy Logic Meets Deep Learning to Crack Handwritten Tamil and Translate It Across 50 Languages

Handwritten Tamil has long been one of the hardest problems in optical character recognition. The script is one of the oldest Dravidian writing systems still in daily use, and its cursive curves, diacritical marks, and ornamental flourishes make even the best neural networks stumble. Now a team of researchers at Vellore Institute of Technology in Chennai has unveiled a pipeline that tackles the problem with an unusual weapon: fuzzy logic. Their framework, called THCRT, couples a fuzzy inference system with deep convolutional recognition and a multilingual transformer, achieving 98.23 percent recognition accuracy and a character error rate of just 1.12 percent on a standard benchmark of handwritten Tamil words.

The core insight behind the work is that most end-to-end recognition pipelines treat handwriting as if it were printed text. They segment the image, extract features, and classify, but they never explicitly model the structural uncertainty that comes from the way different people write. In Tamil, this uncertainty is acute. Decorative strokes vary enormously in width and shape from writer to writer, and when a recognition model misreads a character, that error propagates straight into downstream tasks like machine translation. The researchers argue that this omission is a fundamental flaw in existing architectures, and their solution is to insert a fuzzy uncertainty gate between segmentation and feature extraction.

The gate is called BEMFFIS, short for Biparametric Exponential Membership Function-based Fuzzy Inference System. It works by measuring the coefficient of variation of stroke widths within each segmented character, a statistic that captures how much the decorative elements fluctuate. Characters whose stroke widths are already consistent pass through with minimal normalization, while those with high variability undergo full fuzzy normalization. The clever part is the membership function itself: by coupling two parameters into a single degree of freedom, the researchers dramatically shrink the hyperparameter search space, reducing the risk of overfitting on a small validation set while still adapting to the quirks of individual handwriting styles.

Before any of this can happen, the pipeline must segment the handwritten page into lines, words, and characters. Standard connected component analysis in Cartesian coordinates struggles with the curved baselines of Dravidian scripts, so the team developed CPFCA, which applies a polar transformation first. By converting the image to polar coordinates centered on each text line’s foreground pixels, the curved baseline becomes effectively linear, making component analysis far more reliable. The method achieves a segmentation rate of 98.7 percent with an intersection over union of 96.3, taking just 12.4 milliseconds per image.

Once the characters are segmented and stroke-normalized, recognition is handled by GAE-DCSSNN, a deep convolutional neural network with two architectural twists. First, it replaces standard global average pooling with a two-layer channel excitation mechanism that learns to weight the most discriminative feature channels more heavily. Second, it uses Swish activation instead of ReLU, which avoids the information loss that ReLU suffers for negative inputs. The network processes both the spatial character images and a rich feature vector that includes Fast Fourier Transform components, Zernike moments, histogram of oriented gradients, and structural features like endpoint and intersection counts.

The final stage is translation. The recognized Tamil words are tokenized using the same SentencePiece tokenizer used by mBART, a multilingual sequence-to-sequence model pre-trained on 50 languages. The team augmented this backbone with a Bernoulli distribution-based stochastic token-pruning scheme during training, which acts as structured dropout and forces the encoder to focus on semantically dense embeddings. They also added a dependency parsing step that assigns each subword a role as either head or dependent, reflecting the root-plus-affix morphology of Tamil. During inference, all pruning masks are disabled, ensuring deterministic output.

The evaluation was conducted on two publicly available datasets: the Tamil Offline Handwritten Word Recognition dataset with over 103,000 word images across 13,294 word types, and the Tamil Handwritten Character Corpus from Mendeley Data with 24,700 images across 247 character classes. Both used writer-independent splits to prevent the model from simply memorizing individual handwriting styles. Translation quality was assessed on a 1,200-sample parallel reference set annotated by two certified bilingual translators working independently, with an inter-annotator agreement measured by Cohen’s kappa of 0.91.

The ablation studies reveal exactly where the performance gains come from. Removing the fuzzy inference module drops accuracy by 1.79 percentage points and increases the character error rate by a factor of 1.49. Replacing it with a standard Gaussian fuzzy system recovers some but not all of the performance. Swapping the polar-transform segmentation for conventional connected component analysis costs another 0.955 percentage points. On the translation side, removing dependency parsing reduces the BLEU score from 0.438 to 0.391, while removing both the pruning scheme and the parsing step drops it to 0.312, a relative gain of 40.4 percent for the full system.

The researchers also conducted a ground truth quality verification step before finalizing their benchmark splits, with two annotators reviewing samples for mislabels and visual confusions. This verification alone improved accuracy across six different architectures by an average of 2.14 percentage points, highlighting how much benchmark label quality constrains reported performance in this field. The team notes that their 98.23 percent figure is measured on the quality-verified partition, and that comparisons with prior work using custom datasets should be treated as indicative rather than definitive.

The framework has clear applications in document digitization, automated analysis of historical records, and cross-language communication, particularly for low-resource languages where handwritten materials remain the primary source of information. The authors acknowledge limitations: the system does not explicitly model inter-writer stylistic variability within a document, translation is evaluated only on word-level and short-phrase content, and extending the segmentation and fuzzy modules to other Dravidian scripts like Telugu and Kannada remains future work. Still, the demonstration that a principled fuzzy uncertainty model can measurably improve downstream NLP performance offers a template that could generalize well beyond Tamil.

Subject of Research: Fuzzy logic-augmented deep learning for handwritten Tamil character recognition and multilingual translation

Article Title: A fuzzy-augmented deep learning pipeline for multilingual natural language processing from handwritten script: a Tamil case study

Article References: Kumar A, H., Aarthi, B., Kumaran, K., Saranya, G., & Venkatram C, P. (2026). A fuzzy-augmented deep learning pipeline for multilingual natural language processing from handwritten script: a Tamil case study. Frontiers in Artificial Intelligence, 9, Article 1859443. https://doi.org/10.3389/frai.2026.1859443

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1859443

Keywords: Tamil handwriting recognition, fuzzy inference system, deep learning, multilingual translation, natural language processing, convolutional neural network, mBART transformer, Dravidian scripts, character error rate, stroke uncertainty modeling, polar segmentation, machine translation