Word Spacing Equalization for LSTM Text Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing intelligent character recognition (ICR) systems face challenges in word segmentation due to variable word spacings, leading to unsuccessful training of recurrent neural networks (RNN) and inaccurate character recognition.
Innovation Solution
A method that preprocesses text line images by normalizing their height and equalizing word spacings to a standard length, which are then used to train and predict characters using an LSTM neural network, without modifying the network model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If text line images are used for training RNN with variable word spacings, then the system can handle natural text layout, but the training fails due to incorrect space length detection by automatic labeling algorithms
Solution Approach 1:
The patent applies word spacing equalization as a preliminary preprocessing step before feeding text line images to the RNN for training. By pre-adjusting all word spacings to a standard length and pre-determining correct space boundaries, the system eliminates the need for the RNN to learn variable spacing patterns, thereby ensuring reliable training while preserving natural text layout handling
2Measurement precision
If word segmentation is performed to handle variable spaces, then character recognition can be improved, but the segmentation task becomes challenging and complex
Solution Approach 1:
The patent changes the spacing parameter from variable to fixed by equalizing all word spaces to a standard length. This parameter transformation simplifies the segmentation task significantly, as the fixed spacing eliminates the need for complex algorithms to detect variable space boundaries, while still enabling accurate character recognition through the standardized structure
Solution Approach 2:
The patent extracts and standardizes the spacing component separately from the content. By isolating the word spacing element and assigning it a fixed standard length, the system removes the complexity of variable spacing detection while maintaining the structural integrity needed for accurate character recognition
Data Source
AI summary
In an intelligent character recognition (ICR) method for recognizing hand-written text images using a long-short term memory (LSTM) recurrent neural network (RNN), text images are segmented into text line images, and the text lines images are pre-processed to normalize the line height and to equalize the word spacings in each text line. Both training images used to train the RNN network and test images containing text to be recognized by the trained RNN network are pre-processed to have identical heights and identical word spacings between words. This method improves character recognition accuracy.