Document Image Binarization via Phase Congruency Patch Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for binarizing and segmenting document images struggle to accurately distinguish and separate text from non-text content, especially in gray scale images containing both, which is crucial for optical character recognition and document processing.
Innovation Solution
The method calculates phase congruency maps at different orientations, combines them, and uses these maps for texture analysis, edge extraction, and statistical characterization of connected components to segment text and non-text areas, followed by binarization using specific threshold values for each image patch.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional binarization methods are used on gray scale document images, then processing speed is maintained, but the accuracy of distinguishing text from non-text content deteriorates
Solution Approach 1:
The patent divides the document image into multiple image patches and processes each patch individually with orientation-specific phase congruency analysis. This segmentation approach enables accurate text/non-text distinction in each local region while maintaining overall processing efficiency through parallelization.
Solution Approach 2:
The patent introduces orientation as an additional dimension by calculating phase congruency maps at multiple orientations (0°, 45°, 90°, 135°) for each image patch. This multi-orientation analysis captures text structures in different directions, significantly improving text segmentation accuracy compared to single-orientation methods.
2Measurement precision
If phase congruency maps at multiple orientations are calculated and combined, then text segmentation accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent segments the image into patches and processes only relevant patches with multiple orientations. By identifying patches that likely contain text first (through initial analysis) and applying full multi-orientation phase congruency only to those patches, computational power is conserved while maintaining high edge detection accuracy where needed.
Solution Approach 2:
The patent applies different processing intensities to different regions: full multi-orientation phase congruency analysis is applied locally to patches identified as containing text, while simpler methods are used for non-text patches. This local quality approach optimizes the balance between edge detection accuracy and computational power consumption.
3Measurement precision
If connected component analysis is performed on binary edge maps, then text area identification is improved, but processing time increases
Solution Approach 1:
The patent performs connected component analysis on segmented image patches rather than the entire image. This patch-based approach reduces the computational burden of connected component extraction and allows parallel processing of multiple patches, maintaining high text area segmentation precision while reducing overall processing time.
Data Source
AI summary
A method is described for binarizing a gray scale document image, in particular, a document image containing both text and non-text contents. Phase congruency maps are calculated from the gray scale image, and used to segment the text and non-text areas of the gray scale image. The phase congruency maps are also used to extract long lines such as table lines, which can be optionally removed from the image. The text and non-text areas of the gray scale image are divided into image patches; for the text areas, connected components obtained from the phase congruency map are used to generate image patches, so that each image patch contains a text character. The image patches are binarized individually using individual threshold values, and then combined to generate a binary image of the gray scale image. The method can also be used for purposes of OCR or document authentication.