3D De-skewing Digital Image Text Lines
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Digital camera images of text documents often suffer from distortion due to non-flat document surfaces and non-perpendicular camera angles, making them unsuitable for Optical Character Recognition (OCR) as they contain curved and skewed text lines.
Innovation Solution
A method is developed to de-skeew digital images by identifying text, determining its direction, and applying a three-dimensional transformation to straighten text lines, which includes identifying black connected components, determining horizontal and vertical directions, and applying unwarping transformations to correct for perspective distortions and curvature.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If digital camera images are taken at non-perpendicular angles or from non-flat surfaces, then the images capture more of the document content, but the text lines become distorted and curved making them unsuitable for OCR
Solution Approach 1:
The patent applies a three-dimensional transformation to correct perspective distortion by mapping 3D document surface coordinates to 2D image coordinates. This involves determining vanishing points and creating a transformation matrix that accounts for the camera's angle and the document's curvature, effectively adding a dimensional perspective correction step that converts skewed 2D projections back into straight text lines.
Solution Approach 2:
The system performs preliminary text detection and direction determination before applying the distortion correction. By first identifying text regions and their orientations in the distorted image, the system can pre-calculate the appropriate transformation parameters and vanishing points, then apply the correction in an optimized sequence that preserves text integrity while straightening lines.
2Manufacturing precision
If the camera is positioned perpendicular to the document surface, then text lines remain straight for OCR, but less document content can be captured in the image
Solution Approach 1:
The patent enables oblique viewing angles by applying 3D perspective transformation that mathematically corrects the distortion introduced by non-perpendicular camera positioning. This allows the system to capture wider document areas at angles while recovering straight text lines through coordinate transformation, effectively decoupling the trade-off between coverage area and text straightness.
3Area of stationary object
If the document surface is non-flat, then more document areas can be captured, but text lines become curved and distorted
Solution Approach 1:
The patent models the document surface as a 3D curved manifold and applies differential geometry-based transformation to flatten the curved text lines. By detecting the surface curvature and calculating appropriate warping transformations, the system maps 3D curved surface coordinates to 2D flat image coordinates, straightening text lines while preserving the expanded document area coverage.
4Productivity
If traditional OCR processing is applied directly to distorted camera images, then processing is simpler and faster, but recognition accuracy deteriorates due to skew and curvature
Solution Approach 1:
The patent implements preliminary de-skewing and straightening transformations before OCR recognition. By detecting text directions and applying correction transformations in advance, the system prepares optimized input images for OCR that maintain straight text lines, thereby preserving high recognition accuracy while keeping the overall processing pipeline efficient through pre-computed transformation matrices.
Data Source
AI summary
A method of de-skewing a digital image is described. An input camera image is initially received, and text within the input camera image is identified. A text direction of the identified text is determined to determine text lines within the camera image. A three-dimensional de-skewing transformation is determined of the text lines to make the text lines horizontal. Then the de-skewing transformation is applied to the input camera image to form a de-skewed output image. An unwarping transformation may also be applied to the input camera image for straightening text lines that are curved.