Abstract
This paper presents a script-independent "word spotting" system designed to retrieve specific words from document images without the need for full Optical Character Recognition (OCR). By representing word shapes through statistical Moment Functions, the methodology provides a robust solution for multilingual information retrieval across English, Hindi, and Sanskrit. The system demonstrates significant resilience to font variations, scaling, and script-specific challenges like the Devanagari Shirorekha.
Introduction
Traditional Optical Character Recognition (OCR) systems are inherently script-dependent; a model trained for Latin scripts cannot process Devanagari without significant reconfiguration. Furthermore, scripts like Hindi and Sanskrit present unique challenges due to complex character joining and the Shirorekha—the horizontal line that connects letters within a word.
To address these challenges, this research shifts the focus from character recognition to word spotting. By treating words as visual objects rather than sequences of characters, the system can identify visual instances of a query word within a large corpus of handwritten or printed images. The goal is a universal feature extraction method that prioritizes the geometry of a word's shape over language-specific rules.
Methodology
The proposed system operates through two primary stages: Indexing and Matching.
1. Moment-Based Feature Extraction
The core innovation lies in the application of Geometrical Moments. Document images are first segmented into individual word images and normalized to a standard 256x256 pixel size. The system then calculates statistical moments up to the 7th order, resulting in a 30-dimensional feature vector for each word. Unlike structural features such as Gabor filters, these moments are mathematically invariant to translation and scaling, allowing the system to recognize words regardless of their size or position in the original document.
2. Matching and Relevance Feedback
To retrieve matches, the system employs Cosine Similarity to compare the query word's feature vector against the indexed database. Recognizing that initial automated results may contain noise, the authors integrated a manual feedback loop using Rocchio’s Formula. Users can mark results as relevant or irrelevant, allowing the system to refine the query vector and improve precision in subsequent iterations.
Experimental Results and Analysis
The system was evaluated on three distinct datasets: handwritten English (IAMdb and George Washington collection), printed Hindi (Million Book Project), and printed Sanskrit archives.
The "Sweet Spot" of Complexity
A critical finding of the study is the relationship between moment order and retrieval accuracy. While higher-order moments capture more detail, they also become increasingly sensitive to minor noise. The researchers identified that precision peaks at the 7th order; beyond this point, the statistical signatures become too volatile, leading to a decline in performance.
Comparative Performance
The system consistently outperformed traditional benchmarks like GSC (Gradient, Structural, Concavity) features. Notably, the system proved highly effective at overcoming the Shirorekha in Devanagari scripts, as statistical moments analyze the global distribution of pixels rather than local profiles that the horizontal bar might otherwise mask.
| Script | Precision (Before Feedback) | Precision (After Feedback) |
|---|---|---|
| English (Handwritten) | 66.30% | 69.20% |
| Hindi (Printed) | 71.18% | 74.34% |
| Sanskrit (Printed) | 87.88% | 92.33% |
Sanskrit achieved the highest precision, likely due to the longer, more visually distinct word structures in the dataset, which provided more unique statistical signatures for the moment functions to capture.
Conclusion and Future Work
The research demonstrates that moment-based word spotting is a viable, script-independent alternative to traditional OCR for multilingual retrieval. While the system excels with printed text, handwriting remains a challenge due to high intra-author variability. Future work will focus on scaling the methodology to larger corpora and exploring automated relevance feedback mechanisms to reduce the need for human intervention.
Summary updated