image-ocr

Image OCR Skill

Purpose

This skill enables accurate text extraction from image files (JPG, PNG, etc.) using Tesseract OCR via the pytesseract Python library. It is suitable for scanned documents, screenshots, photos of text, receipts, forms, and other visual content containing text.

For structured receipt fields, follow Structured Field Decision Branches below after the initial extraction. Encode the applicable branches in the extraction and update logic before accepting field values; the raw OCR examples are not a complete field-selection algorithm.

When to Use

• Extracting text from scanned documents or photos
• Reading text from screenshots or image captures
• Processing batch image files that contain textual information
• Converting visual documents to machine-readable text
• Extracting structured data from forms, receipts, or tables in images

Required Libraries

The following Python libraries are required:

import pytesseract
from PIL import Image
import json
import os

Input Requirements

• File formats: JPG, JPEG, PNG, WEBP
• Image quality: Minimum 300 DPI recommended for printed text; clear and legible text
• File size: Under 5MB per image (resize if necessary)
• Text language: Specify if non-English to improve accuracy

Output Schema

The raw OCR helper returns JSON conforming to this schema. For a caller-requested spreadsheet, preserve the caller's final schema and keep auxiliary evidence in a separate working file:

{
  "success": true,
  "filename": "example.jpg",
  "extracted_text": "Full raw text extracted from the image...",
  "confidence": "high|medium|low",
  "metadata": {
    "language_detected": "en",
    "text_regions": 3,
    "has_tables": false,
    "has_handwriting": false
  },
  "warnings": [
    "Text partially obscured in bottom-right corner",
    "Low contrast detected in header section"
  ]
}

Field Descriptions

• success: Boolean indicating whether text extraction completed
• filename: Original image filename
• extracted_text: Complete text content in reading order (top-to-bottom, left-to-right)
• confidence: Overall OCR confidence level based on image quality and text clarity
• metadata.language_detected: ISO 639-1 language code
• metadata.text_regions: Number of distinct text blocks identified
• metadata.has_tables: Whether tabular data structures were detected
• metadata.has_handwriting: Whether handwritten text was detected
• warnings: Array of quality issues or potential errors

Code Examples

Basic OCR Extraction

import pytesseract
from PIL import Image

def extract_text_from_image(image_path):
    """Extract text from a single image using Tesseract OCR."""
    img = Image.open(image_path)
    text = pytesseract.image_to_string(img)
    return text.strip()

OCR with Confidence Data

import pytesseract
from PIL import Image

def extract_with_confidence(image_path):
    """Extract text with per-word confidence scores."""
    img = Image.open(image_path)

    # Get detailed OCR data including confidence
    data = pytesseract.image_to_data(img, output_type=pytesseract.Output.DICT)

    words = []
    confidences = []

    for i, word in enumerate(data['text']):
        if word.strip():  # Skip empty strings
            words.append(word)
            confidences.append(data['conf'][i])

    # Calculate average confidence
    avg_confidence = sum(c for c in confidences if c > 0) / len([c for c in confidences if c > 0]) if confidences else 0

    return {
        'text': ' '.join(words),
        'average_confidence': avg_confidence,
        'word_count': len(words)
    }

Full OCR with JSON Output

import pytesseract
from PIL import Image
import json
import os

def ocr_to_json(image_path):
    """Perform OCR and return results as JSON."""
    filename = os.path.basename(image_path)
    warnings = []

    try:
        img = Image.open(image_path)

        # Get detailed OCR data
        data = pytesseract.image_to_data(img, output_type=pytesseract.Output.DICT)

        # Extract text preserving structure
        text = pytesseract.image_to_string(img)

        # Calculate confidence
        confidences = [c for c in data['conf'] if c > 0]
        avg_conf = sum(confidences) / len(confidences) if confidences else 0

        # Determine confidence level
        if avg_conf >= 80:
            confidence = "high"
        elif avg_conf >= 50:
            confidence = "medium"
        else:
            confidence = "low"
            warnings.append(f"Low OCR confidence: {avg_conf:.1f}%")

        # Count text regions (blocks)
        block_nums = set(data['block_num'])
        text_regions = len([b for b in block_nums if b > 0])

        result = {
            "success": True,
            "filename": filename,
            "extracted_text": text.strip(),
            "confidence": confidence,
            "metadata": {
                "language_detected": "en",
                "text_regions": text_regions,
                "has_tables": False,
                "has_handwriting": False
            },
            "warnings": warnings
        }

    except Exception as e:
        result = {
            "success": False,
            "filename": filename,
            "extracted_text": "",
            "confidence": "low",
            "metadata": {
                "language_detected": "unknown",
                "text_regions": 0,
                "has_tables": False,
                "has_handwriting": False
            },
            "warnings": [f"OCR failed: {str(e)}"]
        }

    return result

# Usage
result = ocr_to_json("document.jpg")
print(json.dumps(result, indent=2))

Batch Processing Multiple Images

import pytesseract
from PIL import Image
import json
import os
from pathlib import Path

def process_image_directory(directory_path, output_file):
    """Process all images in a directory and save results."""
    image_extensions = {'.jpg', '.jpeg', '.png', '.webp'}
    results = []

    for file_path in sorted(Path(directory_path).iterdir()):
        if file_path.suffix.lower() in image_extensions:
            result = ocr_to_json(str(file_path))
            results.append(result)
            print(f"Processed: {file_path.name}")

    # Save results
    with open(output_file, 'w') as f:
        json.dump(results, f, indent=2)

    return results

Tesseract Configuration Options

Language Selection

# Specify language (default is English)
text = pytesseract.image_to_string(img, lang='eng')

# Multiple languages
text = pytesseract.image_to_string(img, lang='eng+fra+deu')

Page Segmentation Modes (PSM)

Use --psm to control how Tesseract segments the image:

# PSM 3: Fully automatic page segmentation (default)
text = pytesseract.image_to_string(img, config='--psm 3')

# PSM 4: Assume single column of text
text = pytesseract.image_to_string(img, config='--psm 4')

# PSM 6: Assume uniform block of text
text = pytesseract.image_to_string(img, config='--psm 6')

# PSM 11: Sparse text - find as much text as possible
text = pytesseract.image_to_string(img, config='--psm 11')

Common PSM values:

• 0: Orientation and script detection (OSD) only
• 3: Fully automatic page segmentation (default)
• 4: Single column of text of variable sizes
• 6: Uniform block of text
• 7: Single text line
• 11: Sparse text
• 13: Raw line

Image Preprocessing

For better OCR accuracy, preprocess images:

from PIL import Image, ImageFilter, ImageOps

def preprocess_image(image_path):
    """Preprocess image for better OCR results."""
    img = Image.open(image_path)

    # Convert to grayscale
    img = img.convert('L')

    # Increase contrast
    img = ImageOps.autocontrast(img)

    # Apply slight sharpening
    img = img.filter(ImageFilter.SHARPEN)

    return img

# Use preprocessed image for OCR
img = preprocess_image("document.jpg")
text = pytesseract.image_to_string(img)

Advanced Preprocessing Strategies

For difficult images (low contrast, faded text, dark backgrounds), try multiple preprocessing approaches:

1. Grayscale + Autocontrast - Basic enhancement for most images
2. Inverted - Use ImageOps.invert() for dark backgrounds with light text
3. Scaling - Upscale small images (e.g., 2x or 3x) before OCR to improve character recognition. For faint or thin fonts on receipts, 3x with autocontrast frequently recovers decimal points and punctuation that 1x/2x lose.
4. Thresholding - Convert to binary using img.point(lambda p: 255 if p > threshold else 0) with different threshold values (e.g., 100, 128)
5. Sharpening - Apply ImageFilter.SHARPEN to improve edge clarity

Multi-Pass OCR Strategy

For challenging images, a single OCR pass may miss text. Use multiple passes with different configurations:

1. Try multiple PSM modes - Different page segmentation modes work better for different layouts (e.g., --psm 6 for blocks, --psm 4 for columns, --psm 11 for sparse text)

2. Try multiple preprocessing variants - Run OCR on several preprocessed versions of the same image

3. Retain results - Keep each pass's text and source separately for structured-field decisions. Concatenation can improve raw-text coverage, but it loses the evidence needed to select one field value.

def multi_pass_ocr(image_path):
    """Run OCR with multiple strategies and combine results."""
    img = Image.open(image_path)
    gray = ImageOps.grayscale(img)
    w, h = img.size

    # Generate preprocessing variants (include an upscaled variant; 3x often
    # recovers thin punctuation such as the decimal point in prices)
    variants = [
        ImageOps.autocontrast(gray),
        ImageOps.invert(ImageOps.autocontrast(gray)),
        gray.filter(ImageFilter.SHARPEN),
        ImageOps.autocontrast(img.resize((w * 3, h * 3), Image.LANCZOS).convert('L')),
    ]

    # PSM modes to try
    psm_modes = ['--psm 6', '--psm 4', '--psm 11']

    all_text = []
    for variant in variants:
        for psm in psm_modes:
            try:
                text = pytesseract.image_to_string(variant, config=psm)
                if text.strip():
                    all_text.append(text)
            except Exception:
                pass

    # Combine all extracted text
    return "\n".join(all_text)

This approach improves extraction for receipts, faded documents, and images with varying quality.

Structured Field Decision Branches

After one batch extraction, save provisional results and a separate evidence record containing each candidate's raw text, field label, image region, and OCR source. Evaluate the following triggers for every field, including non-empty and format-valid fields. Select a branch before selecting a value. A clear field can be confirmed directly; OCR frequency, keyword rank, and variant order alone are not confirmation evidence.

Branch A — Date glyph inspection before confirmation

Trigger: an unconfirmed date has 8 in the day or month of any recognized candidate. Check the day/month components of every raw and normalized candidate; exclude the year from this digit-only gate. A standalone 3, or 8 occurring only in the year, does not automatically require viewing. For a matching day or month, actually open this receipt's date image before confirmation even when OCR is high-confidence, variants agree, and the date is calendar-valid. Queue each date once; reuse its recorded inspection when its confirmed value is unchanged.

def date_requires_image_review(*day_month_candidates):
    """Pass parsed (day, month) pairs, never complete dates or year tokens."""
    return any(
        "8" in str(day) or "8" in str(month)
        for day, month in day_month_candidates
    )

Any date still requires this branch when a component has broken strokes, unclear loop closure, contact with a separator, conflicting readings, or unresolved day/month/year order. This includes ambiguities in the year and digit confusions other than 3/8. A changed candidate goes through step 4 rather than silently replacing a confirmed value. A nearby Time line can help locate the block and provide same-font comparison digits.

Execute in this order:

1. Preserve the current raw date and each alternative. Crop the complete Date line, including its label, separators, day, month, year, and margins; retain the neighboring Time line when present. Save the original-pixel crop and a larger nearest-neighbor copy. Use these as the reference; sharpened, thresholded, or smoothed versions can manufacture or erase a connecting stroke.

2. Actually open the saved crop with the image-view tool, such as file_editor(command="view", path="<saved crop path>"). Saving a PNG or printing its OCR text does not complete this step. Viewing another receipt or an unrelated image does not clear this date's gate. Read the entire date from this receipt's actual crop, not just the triggering day/month digit. A queued date requires this actual view even when the candidate seems unambiguous. Check the glyphs:

   Visible feature in the original crop	Interpretation to check
   Two rightward-curving lobes joined on the right, with open gaps on the left	An open-arc digit such as 3; do not treat the lobes as two closed rings
   A real left-hand connecting stroke closes both the upper and lower holes	A closed-loop digit such as 8; verify both closures in the original pixels
   Apparent closure exists only after thresholding, scaling, or isolated noise bridges a gap	Processing artifact; return to the original crop and keep alternatives pending

   Compare the glyph with unambiguous printed digits in the same font, including the neighboring time when useful. The time supplies glyph shapes, not a date value. Inspect all complete date components rather than choosing a digit from its overall dark outline. Neither a global 8 -> 3 replacement nor the reverse is a valid repair.

3. Use focused line OCR only to generate additional candidate readings. Resolve the actual printed glyph using the crop and its label context, then establish day/month/year order and convert to ISO. Write a separate date-evidence record with the accepted value, crop path, visible distinguishing strokes, and status confirmed. If the original pixels cannot distinguish the candidates, retain status pending; format validity, first-variant order, and votes cannot promote it to confirmed.

4. Load that evidence record on every later extraction. A confirmed date is the accepted date for the deliverable, independent of new OCR order or votes. Put any disagreeing new reading into review without overwriting it. Replace the accepted date only after another actual crop inspection supplies stronger same-field evidence, and record the reason. A provisional date is not protected merely because it happened to appear in an earlier workbook.

5. Commit a newly confirmed or corrected date before moving to another receipt: persist the accepted value and evidence, update that receipt's date in the working rows, and actually save the target workbook. If it already exists, update the affected cell while preserving unrelated accepted fields; otherwise create it from the current working rows. Reopen the workbook and read back the changed cell. Editing an override in a script is only preparation: execute the targeted update and verify the saved value now, not after the remaining review queue. This commit needs no new full-batch OCR.

Confirm when: the image-view operation for this receipt's date has occurred, the printed-date reading, relevant stroke evidence, and complete date context are recorded, and workbook readback matches the accepted date. Reuse that recorded inspection for an unchanged confirmed date; a changed reading requires the review in step 4. When no image-view tool is available, acknowledge that limitation and keep a triggered date pending rather than claiming a visual check. Keep the evidence record outside the requested workbook schema.

Branch B — Inclusive total, rounding, and a later final total

Trigger: the payment block contains a pre-round total (possibly labeled as tax-inclusive), a signed rounding adjustment, and a later TOTAL, NET TOTAL, NETT TOTAL, or other final-payable line. Different amounts in these roles are not conflicting OCR readings of one field.

Execute:

1. Crop and read the whole payment block together. Assign candidates to separate roles: subtotal, tax/service components, pre-round payable, signed adjustment, final payable, tendered cash, and change. Keep a tax-summary table outside this payment block; its own TOTAL is not the amount due.
2. Preserve the + or - on the adjustment. Parse these monetary tokens with Decimal and check final_payable == pre_round_payable + signed_adjustment. The sign is part of the value, not punctuation to discard. If the pre-round payable already includes tax/service charges, use those components only to cross-check it, not to add them again.
3. Select the explicitly printed final-payable amount when the field association and arithmetic agree, even if an earlier tax-inclusive amount occurs in more OCR passes or has a higher-ranked keyword. The final label can be larger or bolder than surrounding text; inspect that line separately if OCR misses it.
4. If a digit or separator is missing or the arithmetic disagrees, review the corresponding local crop. Accept the result only after resolving the reading; do not invent a rounding adjustment or silently force the arithmetic to fit.

Confirm when: the final amount is anchored to its own label and verified against the printed settlement relation. If no rounding/final-total pattern is present, resolve the final-payable role from the receipt's actual labels; do not apply this branch's adjustment speculatively. Keywords such as GST identify a possible component, not an automatic reason to discard an inclusive payable.

Branch C — Amount split across lines or repaired digits

Trigger: an amount keyword has no number on its line, or a candidate relies on inserting a decimal, removing an OCR character, or joining separated digits.

Execute: crop the label and neighboring lines together. Locate the numeric region that is visually associated with that label; exclude quantity, date, identifier, tax-rate, and change fields. Use next-line/trailing tokens only as candidates. Support each punctuation/digit repair from the original monetary region, and then apply Branch B if settlement roles are present.

Confirm when: both the label association and repaired number are supported. Neither the last number on a line nor successful parsing clears this branch.

Branch D — Reusing confirmed fields after a parser change

Trigger: changing shared parsing/selection logic, adding a raw-OCR preference, or retrying extraction after any fields have been confirmed.

Execute: save and reload the confirmed field values with their evidence before the change. Run the new logic into provisional candidates. Merge only reviewed updates to unresolved or demonstrably affected fields; retain every other confirmed field. A changed candidate requires stronger same-field evidence and a recorded reason before replacing a confirmed value. Rebuild the deliverable from this merged accepted state, not from fresh unreviewed batch output.

Confirm when: a before/after field comparison shows that every replacement has been reviewed. Execute the write, reopen the destination workbook, and compare its stored values with the accepted state. Logs or values left only in memory do not demonstrate delivery. Reserve budget for this final readback instead of unnecessarily repeating full-batch OCR at the end. Reserve at least the last five available model/tool iterations for delivery: save the merged accepted state, execute any still-unapplied targeted updates, reopen the actual output, and compare its stored fields, coverage, schema, and ordering with that state. Check the remaining budget before starting another OCR experiment or group of image views. Once the delivery reserve is reached, finish writing and readback instead of starting optional rechecks. If required fields remain unresolved, save the current work and report them rather than guessing or claiming completion.

Numeric Normalization for Receipts and Forms

OCR frequently corrupts decimal separators and digit spacing in monetary amounts. Apply the following repairs only inside a monetary region associated with its label, and only when its printed evidence supports the repair (Branch C):

• Replace : between digits with . (e.g., 92:80 -> 92.80).
• Collapse a single space between a 1-4 digit group and a 2-digit group into . (e.g., 92 80 -> 92.80, 1 234 56 -> keep the comma-thousands intact but treat trailing 56 as decimals only when no other decimal separator is present).
• Reinterpret bare digit clusters of length >=3 whose last two digits look like cents as X.YY when no decimal separator was recognized (e.g., 9280 -> 92.80) only when the line contains a strong keyword such as GRAND TOTAL and no better candidate exists.
• Remove clearly extraneous surrounding OCR marks; retain decimal/thousands punctuation and any genuine sign, especially on rounding adjustments.
• Accept commas as thousands separators: 1,234.56 -> 1234.56.

Prefer a well-formed monetary token in the identified field. Next-line candidates and decimal repairs must pass Branch C; this helper parses a token but does not decide whether it is the final amount. Keep date and ID parsing separate.

import re
from decimal import Decimal

MONEY_STRICT = re.compile(r'[+-]?(?:\d{1,3}(?:,\d{3})+|\d+)\.\d{2}')

def parse_money(token: str, *, decimal_repair_supported=False):
    """Parse a field-associated monetary token, preserving a printed sign."""
    if token is None:
        return None
    s = token.strip().replace('\u2212', '-')
    if MONEY_STRICT.fullmatch(s):
        return Decimal(s.replace(',', ''))
    if not decimal_repair_supported:
        return None
    # Enable only after the original region establishes the decimal position.
    candidates = [
        re.sub(r'(\d):(\d{2})$', r'\1.\2', s),
        re.sub(r'(\d)\s+(\d{2})$', r'\1.\2', s),
    ]
    if re.fullmatch(r'[+-]?\d{3,}', s):
        candidates.append(s[:-2] + '.' + s[-2:])
    for candidate in candidates:
        if MONEY_STRICT.fullmatch(candidate):
            return Decimal(candidate.replace(',', ''))
    return None

Error Handling

Common Issues and Solutions

Issue: Tesseract not found

# Verify Tesseract is installed
try:
    pytesseract.get_tesseract_version()
except pytesseract.TesseractNotFoundError:
    print("Tesseract is not installed or not in PATH")

Issue: Poor OCR quality

• Preprocess image (grayscale, contrast, sharpen)
• Use appropriate PSM mode for the document type
• Ensure image resolution is sufficient (300+ DPI)

Issue: Empty or garbage output

• Check if image contains actual text
• Try different PSM modes
• Verify image is not corrupted

Issue: Keyword detected but no numeric value parsed

• Do not immediately return null. First:
  1. Run Branch C on the label and its neighboring numeric region.
  2. If multiple settlement amounts or rounding appear, run Branch B.
  3. Retain unresolved candidates and continue focused review within the available budget.
• If the field remains unresolved, report extraction failure instead of accepting an unrelated but parseable number. Non-empty ambiguous fields use the same decision branches; review is not limited to null values.

Quality Self-Check

Before returning results, verify:

• [ ] Output is valid JSON (use json.loads() to validate)
• [ ] All required fields are present (success, filename, extracted_text, confidence, metadata)
• [ ] Text preserves logical reading order
• [ ] Confidence level reflects actual OCR quality
• [ ] Warnings array includes all detected issues
• [ ] Special characters are properly escaped in JSON

Structured-Value Extraction Self-Check

When the caller extracts specific fields (amounts, dates, IDs), also verify:

• [ ] For every image where a target keyword (e.g., GRAND TOTAL, TOTAL RM) is detected in any OCR variant, a candidate value was produced, or the reason for failure is explicitly logged.
• [ ] Every unconfirmed date with 8 in a candidate day/month passed Branch A with an actual view; year-only 8 and standalone 3 did not trigger unnecessary viewing, while ambiguous or changed dates still received review.
• [ ] Each accepted date correction was saved and read back before continuing the review queue; unchanged confirmed dates reused their recorded inspection.
• [ ] Settlement totals passed Branch B where applicable; pre-round, rounded, tax-summary, tendered, and change amounts were kept in separate roles.
• [ ] Decimal repairs and split-line candidates passed Branch C, preserving a genuine adjustment sign.
• [ ] Parser changes and retries passed Branch D; confirmed fields were preserved unless an evidenced correction was accepted.
• [ ] The delivery reserve was used to apply all accepted updates to the requested file, and readback matched the accepted fields and caller's schema, coverage, and ordering.

Limitations

• Tesseract works best with printed text; handwriting recognition is limited
• Accuracy decreases with decorative fonts, artistic text, or extreme stylization
• Mathematical equations and special notation may not extract accurately
• Redacted or watermarked text cannot be recovered
• Severe image degradation (blur, noise, low resolution) reduces accuracy
• Complex multi-column layouts may require custom PSM configuration
• Thin punctuation (decimal points, colons) is often lost or substituted; always attempt numeric normalization before rejecting a value

Version History

• 1.0.0 (2026-01-13): Initial release with Tesseract/pytesseract OCR
• 1.1.0: Added value-level voting across variants, numeric normalization guidance for OCR-corrupted amounts, and structured-value extraction self-check.