Scene Text Orientation Correction for Vertical and Slanted OCR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing text detection and recognition systems struggle with detecting and recognizing textual regions in non-horizontal orientations, such as vertical or slanted orientations, due to their reliance on traditional OCR algorithms designed primarily for horizontal text.
Innovation Solution
A method and system utilizing a trained deep convolutional network to detect and recognize alpha-numeric characters in various orientations by employing a bounding box estimator, gradient/tilt estimator, image rotator, and optical character reader (OCR) to align and correct text orientations, interpolate missing characters, and perform horizontal warping for accurate recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional OCR algorithms are used, then horizontal text recognition is accurate, but vertical or slanted text detection fails
Solution Approach 1:
The system dynamically adjusts the orientation of text regions by calculating tilt angles and rotating images to align text horizontally. This dynamic adaptation allows the OCR algorithm to maintain high accuracy across various text orientations without requiring multiple specialized algorithms for different orientations.
Solution Approach 2:
The system changes the orientation parameter of text regions by rotating images based on calculated tilt angles. By transforming the orientation parameter from vertical/slanted to horizontal, the system enables traditional OCR algorithms to accurately recognize text regardless of its original orientation.
2Measurement precision
If image rotation is applied to correct text orientation, then vertical text recognition improves, but processing time increases
Solution Approach 1:
The system performs preliminary detection of text regions and their orientations before applying rotation. By identifying only the regions that require rotation and pre-calculating their tilt angles, the system minimizes unnecessary processing and reduces overall processing time while maintaining detection accuracy.
Solution Approach 2:
The system segments the image into multiple text regions and processes each region independently based on its specific orientation. This segmentation allows the system to apply rotation only to regions that need it, rather than rotating the entire image, thereby reducing processing time while maintaining accuracy for oriented text.
3Adaptability or versatility
If multiple processing steps are added for orientation correction, then text recognition in various orientations improves, but system complexity increases
Solution Approach 1:
The system uses a universal approach by rotating images to a standard horizontal orientation, allowing a single OCR algorithm to handle all text orientations. This multi-functional strategy eliminates the need for separate algorithms for different orientations, reducing system complexity while maintaining versatility.
Solution Approach 2:
The system introduces an intermediate orientation correction step that transforms various text orientations into a common horizontal format. This intermediary transformation serves as a bridge between diverse input orientations and the horizontal-text-specialized OCR algorithm, simplifying the overall system architecture.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
Detection of textual phrases in a non-horizontal orientation at a scene is a target problem. This disclosure relates to a processor implemented method to detect a text from multimedia content captured at a scene. An input original image is processed by a trained model to obtain individual character with bounding box on the original image. The original image is positioned by a gradient to obtain a rotated image if number of detected characters is not equal to number of expected characters on the original image. At least one missing character bounding box on the original image and on the rotated image are estimated to construct a horizontal text image if number of detected characters is not equal to number of expected characters on the rotated image. At least one missing character in the estimated bounding box is detected by at least one text returned from an optical character reader.