Text Detection in Natural Images Using Stroke Feature Transform
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current text detection methods in natural scene images face challenges due to the diversity of text patterns and complexity of scenes, leading to difficulties in accurately distinguishing text from non-text regions, especially with variations in font, size, shape, and distortions caused by illumination or occlusion, and false alarms from text-like background objects.
Innovation Solution
A system and method that employs a component detection module using a Stroke Feature Transform (SFT) filter, which generates stroke width and color maps to identify text components, and a component filtering module with Text Covariance Descriptors (TCDs) for robust classification, reducing incorrect connections and enhancing intra-component connections, thereby improving text detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If texture-based methods are used to detect text in images, then text regions can be classified based on extracted window descriptors, but text-lines with layout variation (rotation, perspective distortion, aspect ratio) cannot be well captured
Solution Approach 1:
The patent segments text detection into multiple levels: first detecting text components at pixel level using SFT filter, then grouping components into text lines at component level, and finally classifying text lines at line level using TCDs. This hierarchical segmentation allows each level to specialize in capturing specific text characteristics, thereby handling layout variations while maintaining detection precision.
Solution Approach 2:
The patent introduces multi-level classification dimensions beyond simple texture analysis. By adding component-level grouping and line-level classification with TCDs, the system creates additional dimensional space for capturing text layout variations such as rotation, perspective distortion, and aspect ratio changes that single-level methods miss.
2Measurement precision
If component-based methods with low-level filters are used, then background pixels can be discarded, but filtering is sensitive to image noise and distortions leading to incorrect component grouping
Solution Approach 1:
The patent changes the parameters used for component identification from simple low-level filters to multi-parameter Text Covariance Descriptors that capture stroke width, color, and orientation variations. This parameter expansion makes the system more robust to noise and distortions while maintaining accurate component identification and grouping.
Solution Approach 2:
The patent combines multiple feature types (stroke width, color, orientation) into composite Text Covariance Descriptors, analogous to composite materials. This combination creates a more robust representation that resists degradation from individual feature noise or distortion, improving both identification accuracy and grouping reliability.
3Measurement precision
If multi-level classification is performed at component and text-line levels, then text detection accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent segments the classification task into distinct levels (component level with SFT filter, text-line level with TCDs), allowing each level to use optimized algorithms appropriate to its specific task. This segmentation manages complexity by preventing any single level from becoming overly complex while achieving high overall accuracy through coordinated multi-level classification.
Data Source
AI summary
A system and method of text detection in an image are described. A component detection module applies a filter having a stroke width constraint and a stroke color constraint to an image to identify text stroke pixels in the image and to generate both a first map based on the stroke width constraint and a second map based on the stroke color constraint. A component filtering module has a first classifier and second classifier. The first classifier is applied to both the first map and the second map to generate a third map identifying a component of a text in the image. The second classifier is applied to the third map to generate a fourth map identifying a text line of the text in the image. A text region locator module thresholds the fourth map to identify text regions in the image.