Amharic OCR & Document Intelligence

1. Morphological and Geometric Challenges

The Ethiopic writing system (an abugida or alphasyllabary) encodes consonants and vowels into unified syllographs across seven standard orders:

Order:     1st (ä)   2nd (u)   3rd (i)   4th (a)   5th (e)   6th (ə)   7th (o)
Syllable:    በ         ቡ         ቢ         ባ         ቤ         ብ         ቦ
IPA:       [bɐ]      [bu]      [bi]      [ba]      [be]      [bɨ]      [bo]

Large Inventory, Small Stroke Differences

An Ethiopic OCR model must distinguish over 300 syllabic glyphs, plus numerical ciphers (፩ to ፼) and punctuation delimiters (፡, ።, ፣, ፤).

Vowel changes appear as small variations on the base consonant:

• Foot appendages: Adding a small right-hand horizontal or vertical foot (e.g., በ → ቡ in the 2nd order, ለ → ሉ).
• Shortening/Extension of limbs: In the 3rd order, shortening the right limb and attaching a horizontal cross-stroke (e.g., በ → ቢ).
• Base stroke modifications: Rounding, slanting, or modifying loops (e.g., በ → ቤ in the 5th order, በ → ቦ in the 7th order).

Confusion Set A (Intra-character order shifts):
  በ (1st: [bɐ])  vs.  ቤ (5th: [be])  vs.  ቦ (7th: [bo])
  ነ (1st: [nɐ])  vs.  ኔ (5th: [ne])  vs.  ኖ (7th: [no])

Confusion Set B (Inter-character structural twins):
  ሀ (hä)  vs.  በ (bä)  vs.  ከ (kä)
  አ (ʔä)  vs.  ዐ (ʕä)  vs.  የ (yä)
  ገ (gä)  vs.  ጉ (gu)  vs.  ጎ (go)

In low-resolution scans, paper bleed, ink spread, and photocopy degradation can erase or connect these strokes.

---

2. Document Degradation

1. Vellum Parchment (Branna) Manuscripts: Hand-scribed with carbon-black and cinnabar-red inks on sheepskin or goatskin. Non-uniform ruling, warping, ink oxidation, and historical Ge’ez abbreviations (seraye) challenge line segmentation.
2. 20th Century Letterpress and Typewriters (Negarit Gazeta, 1940s–1980s): Mechanical wear caused ink spread and tilt. Glyphs often touch or fuse, which breaks heuristic character segmentation.
3. Mimeographed & Thermal Invoices: Tax invoices, land registries, and clinic logs printed on thermal or thin paper with uneven toner density.

---

3. Architectures

Analytical Segmentation

Early Amharic OCR systems (Worku, Cowell, Assabie) used horizontal and vertical projection profiles to slice images into lines, words, and character bounding boxes. Multi-limb characters overlap vertically and letters touch, so segmentation errors compounded downstream.

CRNN (CNN + BiLSTM + CTC)

Connectionist Temporal Classification (CTC) removed the need for character segmentation. The model scans full text lines:

• A Convolutional Network (ResNet / VGG) extracts visual feature maps.
• A Bidirectional LSTM models horizontal sequence dependencies.
• A CTC Loss Layer computes the alignment between visual frames and the target text sequence.

Vision-Language Transformers (TrOCR)

Microsoft’s TrOCR combines a pre-trained Vision Transformer (ViT) image encoder with a language model decoder (e.g., RoBERTa or BART).

flowchart LR
    A["Raw Document Image<br/>(Text Line / Crop)"] --> B["Image Preprocessing<br/>(Binarization & Dewarp)"]
    B --> C["Vision Transformer (ViT)<br/>Patch Embeddings (16x16)"]
    C --> D["Cross-Attention Layer"]
    D --> E["Autoregressive Text Decoder<br/>(Ethiopic Subword BPE)"]
    E --> F["Predicted Unicode Text<br/>'ኢትዮጵያ በታሪክ...'"]

The ViT processes image patches without horizontal slicing, and the decoder’s language model prior can use context to reconstruct smudged characters.

---

4. Implementation: Preprocessing and Tesseract

import cv2
import numpy as np
import pytesseract

class EthiopicDocumentOCR:
    def __init__(self):
        # Use Tesseract 5 LSTM engine with Amharic language pack
        self.tess_config = r'--oem 1 --psm 6 -l amh'

    def preprocess_document(self, image_path: str) -> np.ndarray:
        """Removes paper stains, corrects illumination, and binarizes degraded documents."""
        img = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
        
        # Step 1: Remove illumination gradients via morphological closing
        kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (25, 25))
        background = cv2.morphologyEx(img, cv2.MORPH_CLOSE, kernel)
        normalized = cv2.divide(img, background, scale=255)
        
        # Step 2: Otsu's adaptive thresholding with Gaussian blur
        blurred = cv2.GaussianBlur(normalized, (3, 3), 0)
        _, binarized = cv2.threshold(blurred, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
        
        return binarized

    def recognize_text(self, processed_img: np.ndarray) -> str:
        return pytesseract.image_to_string(processed_img, config=self.tess_config)

# Demonstration
ocr_engine = EthiopicDocumentOCR()
preprocessed = ocr_engine.preprocess_document("legal_gazette_sample.png")
extracted_text = ocr_engine.recognize_text(preprocessed)
print("Extracted Amharic Text:\n", extracted_text)