Efficient Learning-Free Keyword Spotting.

Area of Science:

• Computer Vision
• Document Image Analysis
• Information Retrieval

Background:

• Keyword spotting in historical handwritten documents is challenging due to variations in writing styles, skew, and translation.
• Existing methods often require extensive training data or struggle with diverse document characteristics.

Purpose of the Study:

• To propose a segmentation-based, learning-free Query by Example (QbE) keyword spotting method for handwritten documents.
• To address critical variations in text images and reduce storage requirements for document collections.

Main Methods:

• A three-step process: preprocessing (contrast normalization, main-zone detection), feature extraction (zoning scheme, modified Projections of Oriented Gradients), and matching (Selective Matching algorithm).
• An uneven zoning scheme is employed, with denser zoning on query images for detailed representation and reduced storage.
• The Selective Matching algorithm efficiently computes distances and handles augmented image sets.

Main Results:

• The proposed method effectively overcomes challenges posed by text image variations.
• Significant reduction in storage requirements for document collections was achieved.
• Experimental results on seven public datasets demonstrate superior performance compared to state-of-the-art learning-free techniques.

Conclusions:

• The developed learning-free QbE method offers a robust and efficient solution for keyword spotting in handwritten documents.
• The novel feature descriptor and matching algorithm contribute to improved accuracy and reduced computational load.
• This approach has the potential to enhance accessibility and searchability of historical document archives.