Language Identification in Degraded and Distorted Document Images

Shijian Lu, Chew Lim Tan, Weihua HuangDocument Analysis Systems 2006 · DOI

Abstract

Identifying the language of a document is a critical first step in automated document processing. However, traditional Optical Character Recognition (OCR) methods often fail when images are degraded by noise, low resolution, or skew. This paper presents a robust "Word Shape Coding" (WSC) technique that identifies Latin-based languages by analyzing the overall silhouette of words rather than individual characters. By focusing on macro-level features like peaks, valleys, and internal structures, the proposed system achieves over 99% accuracy on noisy and distorted images, maintaining high performance even on short text snippets.

The Challenge: Beyond OCR

Most language identification systems rely on character-level analysis. While effective for high-quality digital text, these methods struggle with real-world "degraded" images, such as low-quality scans or smartphone photos. The primary obstacles include:

  1. Segmentation Failures: In poor-quality images, characters often merge or break, preventing OCR from identifying individual letters.
  2. Noise Interference: Small linguistic features, such as the dot on an "i" or diacritics (accents), are frequently indistinguishable from random image noise.
  3. Geometric Distortion: Image skew or tilt can disrupt algorithms that require perfectly horizontal text lines.

To address these, the authors developed a method that bypasses character recognition entirely, looking instead at the "shape" of the word.

Methodology: Word Shape Coding (WSC)

The core of the proposed approach is the Word Shape Coding (WSC) scheme. This method characterizes words based on their geometric profile rather than their constituent letters.

Feature Extraction

The system extracts two primary types of features:

  • Local Extremum Points: By tracing the upper and lower boundaries of a word, the algorithm identifies "peaks" (local maxima) and "valleys" (local minima). These points are categorized into three vertical zones: ascenders (e.g., 'h', 't'), the main body (between the x-line and baseline), and descenders (e.g., 'g', 'p').
  • Horizontal Intersection Number: A line is drawn through the middle of the text line, and the number of times it intersects character strokes is recorded.
Figure 4: Feature extraction using local extremum points and the middle line on straight (a) and skewed (b) text.

Preprocessing and Robustness

A key innovation is the intentional removal of small components through "size filtering." While counterintuitive, removing dots on 'i's and accent marks eliminates noise and creates a more stable "macro" shape for the word. This makes the system inherently robust to "touching" characters; since the method analyzes the entire word boundary, merged letters do not significantly alter the resulting code.

Figure 6: Preprocessing step: original text (a) vs. text after size filtering (b), which removes noise and small features like dots and punctuation.

Experimental Results

The researchers evaluated the WSC method using 800 images across eight Latin-based languages (English, French, German, Italian, Spanish, Portuguese, Swedish, and Norwegian).

  • High Accuracy: The system achieved an average identification rate of over 99% for both standard and low-resolution documents.
  • Skew Tolerance: The method maintained 97% accuracy even with a 20-degree tilt, without requiring a separate de-skewing step.
  • Short Text Performance: On snippets containing as few as 20 words, the system achieved 94% accuracy. This was made possible by switching from frequency-based metrics (Bray Curtis distance) to presence/absence metrics (Hamming distance) for short samples.

Conclusion and Future Work

The Word Shape Coding approach provides a highly reliable alternative to OCR-based language identification for degraded documents. While currently optimized for Latin-based scripts, future research aims to adapt these geometric principles to other scripts, such as Chinese or Arabic, and to further improve performance under extreme skew conditions where character boundaries begin to overlap vertically.

Summary updated

BibTeX
@inproceedings{lu2006language,
  author = {Shijian Lu and Chew Lim Tan and Weihua Huang},
  title = {{Language Identification in Degraded and Distorted Document Images}},
  booktitle = {Document Analysis Systems},
  year = {2006},
  doi = {10.1007/11669487\_21},
}
Figure