Document Image Segmentation via Iterative Pixel Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document image segmentation methods are inefficient and time-consuming, particularly in classifying and segmenting 2D objects into distinct content types such as text, graphics, and images, as they often require complex processing steps and struggle to accurately segment mixed content types.
Innovation Solution
A method involving the initialization of a segmentation map, creation of a working map, and iterative refinement using combined maps to segment document images into areas with single types of content, where pixel types are evaluated and reassigned based on neighboring pixels, ultimately grouping background pixels with the most common neighboring type.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex processing steps are used to classify and segment 2D objects into different content types, then segmentation accuracy is improved, but processing time increases significantly
Solution Approach 1:
The patent divides the document image into multiple pixel blocks and processes each block independently through iterative evaluation. This segmentation approach allows parallel processing of different regions, reducing overall processing time while maintaining accurate classification of text, graphics, and image content through localized analysis of pixel characteristics and neighboring pixel relationships.
2Reliability
If traditional object classification methods are used to segment document images, then content type identification is achieved, but the processing complexity increases and speed decreases
Solution Approach 1:
The patent transforms the segmentation problem from complex object-level classification to simpler pixel-level parameter analysis. By evaluating pixel types based on local characteristics and neighboring pixel relationships within defined blocks, the system achieves reliable content type identification through straightforward parameter comparison rather than complex classification algorithms.
Solution Approach 2:
The patent applies iterative processing where pixel blocks are evaluated multiple times with progressively refined classifications. In each iteration, the system processes only the necessary pixel blocks that require reclassification, rather than reprocessing the entire image. This partial action approach maintains high identification reliability while significantly reducing overall processing complexity and time.
3Adaptability or versatility
If existing segmentation methods are used, then document images can be segmented into content types, but the methods struggle with mixed content types and are time-consuming
Solution Approach 1:
The patent applies different processing strategies to different pixel blocks based on their local characteristics. Each pixel block is evaluated independently, allowing the system to adapt to mixed content types within the same image by analyzing local pixel type distributions and neighboring relationships. This localized approach enables versatile handling of text, graphics, and image regions while maintaining high processing speed through parallel block evaluation.
Data Source
AI summary
In a document image segmentation method, pixels of the image are classified into different types such as background, text, table, etc., to generate an initial segmentation map. The initial segmentation map is processed multiple rounds. In each round, a working map is divided into 2×2 pixel blocks; based on pixel types in the block, a corresponding pixel in a combined map is assigned a type, and pixels in a corresponding block in the segmentation map are modified either to change some background pixels to other types or keep them unchanged. The initial segmentation map is used as the working map in the first round, and the combined map of the last round is used as the working map for the next round. After a number of rounds, remaining background pixels of the segmentation map are changed to other types based on the types of their neighboring areas.