Both off-line (document images) and on-line HTR (tablet or e-pen signals) are considered. No prior character or word segmentation is needed. Technology relies on character-level optical models based on Convolutional-Recurrent Neural Networks and Hidden Markov Models, along with Finite-State Lexical and N-Gram Language Models. After model training, for each given text line image, a holistic (“Viterbi”) search provides both an optimal transcription and the corresponding word and character segmentations. Applications: Transcription of ancient and legacy documents, transcription of unconstrained handwritten text in survey forms, etc.
- Transcription of ancient documents
- Indexing & Search in large collections of handwriting images
- Multimodal Interactive Handwritten Text Recognition
- Recognition of mathematical expressions