Pixel-Level Noise Detection in Document Image Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current document analysis systems face challenges in accurately segmenting scanned images due to noise artifacts like halftone, bleeding, and texture noise, which affect the accuracy of image type classification and OCR results, especially in color documents with diverse color combinations.
Innovation Solution
A method for pixel-level noise type classification during pixel color segmentation in a non-overlapped block raster order, which detects and reduces halftone, bleeding, and texture noise by determining dominant colors, applying thresholds, and re-quantizing pixels, thereby improving text segmentation and printing quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a moving window with small size is used for halftone detection, then hardware implementation efficiency is improved, but detection accuracy deteriorates due to lack of context information
Solution Approach 1:
The image is divided into non-overlapped blocks for processing. Each block is independently classified into image types (halftone, contour, or text) based on pixel color analysis. This segmentation allows efficient block-level processing while maintaining adequate context within each block, resolving the contradiction between processing efficiency and detection accuracy.
Solution Approach 2:
The patent changes the processing unit from small moving windows to larger non-overlapped blocks, and changes the classification granularity from block-level to pixel-level within blocks. This parameter change enables both efficient processing and accurate detection by analyzing pixel color values and transitions within each block to identify halftone patterns.
2Productivity
If geometrically coarse block classification is used, then processing speed is improved, but text segmentation accuracy deteriorates due to blurring
Solution Approach 1:
The patent implements two-level segmentation: first dividing the image into non-overlapped blocks for efficient processing, then further segmenting each block into pixel-level color regions. This hierarchical segmentation enables fast block-level classification while maintaining pixel-level precision for text segmentation, avoiding the blurring effect of coarse classification.
Solution Approach 2:
The patent applies different processing strategies to different regions within blocks based on local pixel characteristics. Pixel color analysis is performed at the pixel level within each block, allowing crisp text segmentation in text regions while applying appropriate filtering only in halftone regions, thus maintaining both speed and accuracy.
3Measurement precision
If color pixel segmentation is used for diverse color documents, then representation accuracy is improved, but noise artifacts worsen due to degradation from printing and scanning processes
Solution Approach 1:
The patent performs preliminary noise detection and classification during the pixel color segmentation process itself. By analyzing pixel color values and transitions within blocks before final segmentation, the system identifies and flags noise artifacts (halftone, bleeding, texture) early, allowing subsequent processing to account for these artifacts and improve overall accuracy.
Solution Approach 2:
The patent converts the presence of noise artifacts into beneficial information by using them as classification criteria. Halftone regions, bleeding artifacts, and texture patterns are detected and classified as distinct image types. This conversion allows the system to leverage the artifacts' characteristic patterns for accurate region identification, turning harmful factors into useful classification features.
Data Source
AI summary
A method detects noise in a document image. The image includes pixels representing at least text and image data forming the document. The noise includes at least one of halftone, bleeding and texture noise. The method partitions the image into tiles, formed of pixels, determines a set of dominant colors for each tile, associates each pixel with one of the determined dominant colors, and then determines for each dominant color a fragmentation statistic, and applies a threshold to each statistic to detect the noise in the corresponding dominant color in the tile. Detected noise in the image is reduced by (a) merging colors with noise to one color without noise if the noise is halftone noise or texture noise; and (b) removing the colors with noise if the noise is bleeding noise. The pixels of the tile are then quantized to those remaining dominant colors of the set.