Text segmentation for MRC document compression.

Area of Science:

• Computer Science
• Image Processing
• Data Compression

Background:

• Mixed Raster Content (MRC) standard (ITU-T T.44) offers improved compression/quality tradeoffs for documents.
• Effective document compression relies heavily on accurate foreground/background layer segmentation via binary masks.
• Existing segmentation algorithms impact MRC encoder performance, necessitating advanced solutions.

Purpose of the Study:

• To propose a novel multiscale segmentation scheme for enhanced MRC document encoding.
• To improve the accuracy of text detection and reduce false positives in document segmentation.
• To enhance the quality and reduce the bit rate of decoded documents.

Main Methods:

• Developed a sequential segmentation scheme combining Cost Optimized Segmentation (COS) and Connected Component Classification (CCC).
• COS utilizes a blockwise approach within a global cost optimization framework.
• CCC refines segmentation using a Markov Random Field (MRF) model for connected component feature vectors.
• Integrated COS/CCC into a multiscale framework to handle text of varying sizes.

Main Results:

• The proposed COS/CCC multiscale algorithm demonstrated superior text detection accuracy compared to state-of-the-art methods.
• Achieved a lower false detection rate for non-text features.
• Showcased improved decoded document quality and reduced bit rates.

Conclusions:

• The novel multiscale segmentation scheme significantly enhances MRC document encoding performance.
• The COS/CCC approach offers a robust solution for accurate document segmentation.
• This method provides a better compression/quality tradeoff for digital documents.