Related Experiment Videos
A hybrid neural network model in handwritten word recognition
1Department of Information Management, Chaoyang University of Technology, Taichung, Taiwan.
Summary
A novel hybrid neural network improves handwritten word recognition by accurately identifying characters and reducing errors. This system enhances performance compared to existing methods on standard datasets.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Handwritten word recognition systems require accurate character segmentation and classification.
- Existing systems struggle with ambiguities between character classes and non-character segments.
Purpose of the Study:
- To develop and evaluate a hybrid neural network model for improved handwritten word recognition.
- To create a module capable of assigning accurate character class confidences and handling segmentation errors.
Main Methods:
- A cascaded hybrid neural network combining a Self-Organizing Feature Map (SOFM), gradient descent learning for allograph mapping, and a Multi-Layer Feedforward Network (MLFN).
- The system processes image segments to assign confidence values for character classes, addressing ambiguities and segmentation inaccuracies.
Main Results:
- The proposed hybrid neural network system demonstrated superior performance compared to a baseline system.
- Experiments conducted on a standard test set from the SUNY/USPS Database validated the system's effectiveness.
Conclusions:
- The developed hybrid neural network offers a robust solution for handwritten word recognition.
- The system effectively handles character ambiguities and segmentation errors, leading to improved recognition accuracy.