Single-Character Detection Training for Touching and Fragmented Text
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current text detection datasets lack sufficient annotations at a single-character level, leading to insufficient accuracy in single character detection models, particularly for English texts with touching characters and Chinese texts with false radical detection.
Innovation Solution
A method involving a multi-scale feature fusion residual network is used to train a single character detection model with a synthesis text image set and a real scenario text image set, where synthesis text images are generated by combining real scenario backgrounds with random words, and pseudo labels are created from line text annotations to enhance training accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a line text detection algorithm is used to locate single character regions, then line text detection can be achieved, but the positioning accuracy of single characters is insufficient
Solution Approach 1:
The patent segments the line text detection task into two independent sub-tasks: (1) detecting text regions using a text region detection network, and (2) detecting individual characters within those regions using a character detection network. This segmentation allows each network to be optimized for its specific task, with the character detection network focusing exclusively on precise character localization rather than also performing line-level detection.
Solution Approach 2:
The patent introduces text region annotation boxes as an intermediary element that bridges line text detection and single character detection. The text region detection network outputs text region boxes, which then serve as input constraints for the character detection network to search for characters within those regions. This intermediary structure enables the system to achieve both line text detection and precise character positioning.
2Ease of manufacture
If a single character detection model is trained with limited single-character level annotations, then training can proceed, but the detection accuracy is insufficient
Solution Approach 1:
The patent performs preliminary action by first training a text region detection network to identify text regions in images. Once text regions are detected and annotated, these regions are extracted and used as training data for the character detection network. This preliminary detection and annotation step creates a foundation of labeled data that enables subsequent character-level training without requiring manual annotation of every character from scratch.
Solution Approach 2:
The patent uses copying by replicating and reusing the text region annotation data for character detection training. The text region boxes detected by the first network are copied and used as training annotations for the character detection network, allowing the system to leverage existing detection results as training data for the next stage without requiring separate manual annotation efforts.
3Productivity
If English text detection is performed with traditional methods, then detection can be achieved, but touching characters are easily misdetected
Solution Approach 1:
The patent applies local quality by training the character detection network to focus on local character features within text regions. The network learns to distinguish individual character boundaries by analyzing local patterns and features within the text region, rather than relying on global text structure. This local-focused approach enables the model to accurately separate touching characters by identifying subtle local boundary features.
4Productivity
If Chinese text detection is performed with traditional methods, then detection can be achieved, but false detection of radicals occurs
Solution Approach 1:
The patent applies local quality by training the character detection network to recognize complete character structures through local feature analysis. The network learns to identify the full character boundary by analyzing local patterns, preventing it from stopping at radical boundaries. This approach ensures that entire Chinese characters are detected as complete units rather than being fragmented into radicals.
Data Source
AI summary
A training method includes: obtaining a synthesis text image set, a synthesis text image being obtained through synthesizing a real scenario background image and a random word, the synthesis text image being provided with a line text annotation box and a single character annotation box; training an initial algorithm network using the synthesis text image set, so as to obtain an intermediate model; processing a real scenario text image set using the intermediate model, so as to obtain a pseudo label of a real scenario text image, the real scenario text image being provided with a line text annotation box, the pseudo label being a single character annotation box; and training the intermediate model using the synthesis text image set and the real scenario text image set having the pseudo label, so as to obtain the single character detection model.