Handwriting Recognition Data Normalization via RNN Height Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current handwriting recognition technologies face challenges in achieving high accuracy due to inadequate data normalization methods, particularly for cursive handwriting, as existing methods focus on character-level normalization rather than global scales, resulting in low recognition accuracy rates.
Innovation Solution
A method that optimizes data normalization by selecting the best height normalization setting using Recurrent Neural Networks (RNNs) with datasets of multiple sample images, involving the calculation of top place ratios and validation of recognition results to generate normalized images for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional character-level normalization methods are used, then the normalization process is simple and fast, but the handwriting recognition accuracy remains low (no more than 60% for cursive handwriting)
Solution Approach 1:
The patent transitions from character-level normalization to sentence-level normalization, changing the scale dimension from local to global. This allows the system to capture contextual relationships across entire sentences rather than isolating individual characters, thereby improving recognition accuracy for cursive handwriting while managing the increased complexity through systematic processing steps.
Solution Approach 2:
The patent introduces dynamic parameter adjustment by training RNN models with different normalization settings and selecting the optimal configuration based on validation performance. This allows the system to adapt normalization parameters (such as scaling factors and transformation types) to specific datasets and handwriting styles, improving accuracy without requiring a completely complex fixed system.
2Measurement precision
If global-scale normalization is applied to entire sentences or paragraphs, then the handwriting recognition accuracy improves, but the processing time and computational resources increase
Solution Approach 1:
The patent performs preliminary normalization processing on training datasets before model training, including resizing, cropping, and deskewing operations. By pre-processing the data in advance and storing normalized versions, the system reduces real-time processing requirements during actual recognition tasks, balancing accuracy improvement with time efficiency.
Solution Approach 2:
The patent replaces traditional mechanical normalization techniques with deep learning-based RNN models that automatically learn optimal normalization transformations. The RNN models can perform normalization through learned transformations rather than explicit geometric operations, reducing computational overhead while maintaining or improving accuracy.
3Measurement precision
If multiple normalization settings are evaluated through RNN training, then the optimal normalization parameter is identified for better accuracy, but the training time and computational cost increase
Solution Approach 1:
The patent implements self-service optimization where the RNN model automatically evaluates different normalization settings and selects the optimal configuration through validation performance. The system performs self-assessment and self-adjustment of normalization parameters without requiring extensive manual tuning, improving development efficiency while achieving optimal accuracy.
Solution Approach 2:
The patent incorporates feedback loops where normalization settings are evaluated through validation datasets, and the results are used to refine and select the optimal configuration. This iterative feedback process allows the system to converge on the best normalization parameters efficiently, balancing the need for thorough evaluation with development time constraints.
Data Source
AI summary
Disclosed herein is a method of optimizing data normalization by selecting the best height normalization setting from training RNN (Recurrent Neural Network) with one or more datasets comprising multiple sample images of handwriting data, which comprises estimating a few top place ratios for normalization by minimizing a cost function for any given sample image in the training dataset, and further, determining the best ratio from the top place ratios by validating the recognition results of sample images with each top place ratio.