A Pix2Pix Architecture for Complete Offline Handwritten Text Normalization

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Offline handwritten text recognition relies on preprocessing normalization algorithms.
• Existing methods often use heuristic strategies not integrated with recognition models.
• This limits the unified training of normalization and recognition components.

Purpose of the Study:

• To introduce a Pix2Pix trainable model for normalizing handwritten text images.
• To enable seamless integration of normalization as the initial stage in deep learning recognition architectures.
• To facilitate unified training of normalization and recognition while maintaining module interpretability.

Main Methods:

• Utilized a Pix2Pix conditional generative adversarial network for image normalization.
• Trained the model on a blend of heuristic transformations to address handwriting variability.
• Integrated the normalization approach as the first step in a deep recognition architecture.

Main Results:

• Achieved slope and slant normalization, and normalized ascender/descender sizes.
• The proposed method replicated and, in some cases, surpassed a widely used heuristic algorithm.
• Demonstrated effectiveness across two metrics when integrated into a deep recognition architecture.

Conclusions:

• The Pix2Pix model offers an effective, integrated approach to handwritten text normalization.
• This method mitigates intra-personal handwriting variability, enhancing recognition performance.
• The trainable normalization is a viable alternative to traditional heuristic preprocessing techniques.