A novel word spotting method based on recurrent neural networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Keyword spotting is crucial for information retrieval from documents.
• Existing methods often require keywords in training data or are less effective for unconstrained handwriting.
• Handwritten document analysis presents unique challenges due to variability.

Purpose of the Study:

• To develop a novel, template-free keyword spotting method for handwritten documents.
• To leverage neural networks for improved accuracy and flexibility in keyword retrieval.
• To compare the proposed method against established techniques.

Main Methods:

• A recurrent neural network-based system for unconstrained handwriting recognition was adapted.
• A modified CTC Token Passing algorithm was employed for keyword spotting.
• The system was evaluated against dynamic time warping and hidden Markov model-based approaches.

Main Results:

• The proposed neural network-based keyword spotting method demonstrated superior performance.
• It outperformed both dynamic time warping and hidden Markov model systems.
• Analysis confirmed the advantages of keyword spotting over traditional text line recognition.

Conclusions:

• The novel template-free keyword spotting method is highly effective for handwritten documents.
• Recurrent neural networks combined with CTC Token Passing offer a robust solution.
• This approach advances the field of information retrieval from historical and challenging document collections.