A novel connectionist system for unconstrained handwriting recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unconstrained handwritten text recognition faces challenges due to character segmentation difficulties and the need for contextual information, leading to suboptimal performance in current systems.
• Existing systems predominantly utilize hidden Markov models (HMMs), despite known limitations in handling complex sequence labeling tasks inherent in handwriting recognition.

Purpose of the Study:

• To propose and evaluate a novel recurrent neural network (RNN) architecture for unconstrained handwriting recognition.
• To address the limitations of HMMs in sequence labeling tasks characterized by difficult segmentation and long-range bidirectional dependencies.

Main Methods:

• Developed a specialized recurrent neural network designed for sequence labeling with challenging segmentation and bidirectional context.
• Conducted experiments on two extensive unconstrained handwriting databases (online and offline data).
• Compared the performance of the proposed RNN against a state-of-the-art HMM-based system.

Main Results:

• Achieved word recognition accuracies of 79.7% for online data and 74.1% for offline data.
• Demonstrated significantly superior performance compared to the HMM-based system.
• Validated the network's robustness to varying lexicon sizes and analyzed the contribution of its hidden layers and context utilization.

Conclusions:

• The proposed recurrent neural network offers a superior alternative to traditional HMMs for unconstrained handwriting recognition.
• The novel RNN architecture effectively handles the complexities of character segmentation and bidirectional context, leading to improved accuracy.
• Further research into RNNs for sequence labeling tasks in pattern recognition is warranted.