Recursive Word Segmentation for Variable Spacing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing word segmentation methods fail to accurately segment text lines with significantly variable word spacings, often resulting in inadequate segmentation due to non-uniform spacing distributions and mixed font sizes.
Innovation Solution
A recursive word segmentation method that classifies spacing segments into multiple groups based on length, with additional rounds of classification performed if initial conditions indicate inadequate segmentation, such as bimodal distribution, excessive spread, or presence of long words and few words, to refine the segmentation and achieve accurate word separation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a simple clustering method (e.g., k-means with pre-set two clusters) is used to classify spacing segments, then the device complexity is reduced, but the segmentation accuracy deteriorates when word spacings are significantly variable
Solution Approach 1:
The patent applies segmentation by dividing the classification process into multiple rounds. The first round separates spacing segments into two groups (character spacing and word spacing candidates), and the second round further divides the first group into new subgroups. This multi-stage segmentation approach improves accuracy for variable spacing without requiring overly complex algorithms.
Solution Approach 2:
The patent implements dynamic classification by determining whether to perform a second classification round based on detected conditions (bimodal distribution, excessive spread, or presence of long words with few words). This dynamic adaptation allows the system to increase complexity only when necessary, balancing accuracy requirements with computational efficiency.
2Measurement precision
If additional classification rounds are performed to handle variable spacing, then segmentation accuracy is improved, but processing time increases
Solution Approach 1:
The system dynamically determines whether to perform additional classification rounds based on detected conditions in the spacing segments. By only executing the second classification round when specific conditions are met (bimodal distribution, excessive spread, or long words with few words), the patent avoids unnecessary processing time while maintaining accuracy when needed.
Solution Approach 2:
The patent uses feedback mechanisms by analyzing the results of the first classification round and using detected conditions (bimodal distribution, spread metrics, word length distribution) to determine whether additional processing is required. This feedback loop ensures that extra processing time is invested only when the initial classification indicates potential segmentation issues.
Data Source
AI summary
A word segmentation method uses a recursive technique to segment a text line image into word segments. Spacing segments of the line are obtained; an initial word segmentation is performed to classify the spacing segments based on their lengths into candidate character spacing segments and candidate word spacing segments. The initial segmentation result is evaluated to determine whether the candidate character spacing segments still have a bimodal or multi-modal distribution or a large spread in the distribution, or whether the line contains long words and too few words. If the conditions indicate that the initial segmentation is inadequate, another classification step is performed for the candidate character spacing segments to further classify them into new candidate character spacing segments and new candidate word spacing segments. The process is repeated until the word segmentation is deemed adequate based on the evaluation.