Dependence Models for Searching Text in Document Images

Area of Science:

• Computer Science
• Information Retrieval
• Document Image Analysis

Background:

• Existing word spotting methods struggle with arbitrary text search due to reliance on visual similarity and lack of high-quality OCR.
• Searching document images typically requires users to provide sample word images or synthetic queries, limiting usability.

Purpose of the Study:

• To develop a robust framework for searching arbitrary text within document images, overcoming limitations of current word spotting techniques.
• To enable real-time text search capabilities across diverse scripts, including those lacking commercial OCR support.

Main Methods:

• A Markov Random Field (MRF) framework was proposed for document image analysis and text retrieval.
• Dependencies between visual terms and letter bigrams were learned automatically from noisy OCR output for English documents.
• For scripts without commercial OCR (Telugu, Ottoman), models were trained using manually annotated document images.

Main Results:

• The MRF framework demonstrated effectiveness in real-time arbitrary text searching for English, Telugu, and Ottoman scripts.
• Automatic learning of visual term and letter bigram dependencies improved OCR text search accuracy when combined with the proposed method.
• The trained models successfully resolved arbitrary text queries across books, irrespective of font variations, outperforming a BLSTM baseline.

Conclusions:

• The proposed MRF framework offers a powerful solution for arbitrary text searching in document images across multiple scripts.
• This approach significantly enhances document image retrieval, particularly for languages lacking advanced OCR technology.
• The method provides a scalable and accurate solution for searching historical and diverse textual collections.