Handwritten Text Recognition (HTR) has seen significant advancements in languages like English and Chinese. However, for Indic scripts, which are used by hundreds of millions of people, a substantial digital divide persists. Most existing research focuses on isolated word or line recognition, which fails to address the complexities of real-world document digitization. Researchers from IIT Bombay have introduced PLATTER, an end-to-end framework designed to bridge this gap by providing a robust solution for page-level document analysis across diverse Indic languages.
Methodology
PLATTER employs a modular, two-stage pipeline that decouples text detection from recognition. This architecture allows for flexibility and targeted improvements in each stage.
Stage 1: Handwritten Text Detection (HTD)
The first stage utilizes DBNet (Differential Binarization Network), a segmentation-based detector. A key finding of the research is the effectiveness of a language-agnostic model. By training a single model on a diverse mixture of Indic scripts, the researchers achieved performance comparable to specialized, language-specific models. This approach significantly simplifies the deployment of multi-lingual HTR systems.
Stage 2: Handwritten Text Recognition (HTR)
The second stage focuses on recognizing the text within the detected bounding boxes. The PLATTER framework is designed to be plug-and-play, allowing for the integration of various HTR architectures. The researchers benchmarked six models, ranging from traditional Convolutional Recurrent Neural Networks (CRNN) to modern Transformer-based architectures like MASTER and PARSEQ.
Datasets and Data Augmentation
A major hurdle in Indic HTR is the scarcity of high-quality, page-level datasets. To address this, the authors introduced the Corpus of Handwritten Indic Scripts (CHIPS).
CHIPS was created by "stitching" real handwritten words into 1024x1024 page images, simulating realistic layouts with variable word heights and spacing. Additionally, the researchers utilized ScrabbleGAN to generate 5 million synthetic handwritten word images. This synthetic data acted as a catalyst, significantly accelerating model convergence and improving final recognition accuracy. For instance, in Hindi, the inclusion of synthetic data boosted validation accuracy from 74% to nearly 78% while reducing training time by two-thirds.
Experimental Results
The evaluation of PLATTER yielded several key insights:
- Detection Robustness: The language-agnostic DBNet achieved an impressive F1-score of 0.98 (at IoU 0.5) on the CHIPS test set, demonstrating its ability to learn the "visual signature" of handwriting across different scripts.
- Recognition Performance: In isolated word recognition tasks, PARSEQ emerged as a top performer, achieving a Character Recognition Rate (CRR) exceeding 95% across the board.
- End-to-End Challenges: When evaluating the full pipeline (detection followed by recognition), the Word Recognition Rate (WRR) reached 43.67% using the MASTER model. While this provides a strong baseline, it highlights the significant impact of "detection noise"—imperfect bounding boxes—on final OCR output.
Conclusion and Future Directions
PLATTER represents a significant step toward comprehensive digitization of Indic handwritten documents. By proving the viability of language-agnostic detection and providing the CHIPS dataset, the researchers have laid a foundation for future work. Key areas for improvement include integrating automatic script identification to make the system fully autonomous and developing "detection-aware" HTR models to close the performance gap in end-to-end scenarios. The project's code and models have been made publicly available to encourage further community-driven innovation in Indic language technology.
Summary updated