tocmo0nlord/odoo-ai
1
0
Fork 0
Code Issues 1 Pull Requests Actions Packages Projects Releases Wiki Activity

Improve OCR preprocessing and amount extraction robustness

Browse Source 
Image preprocessing (receipt_parser.py):
- Add ImageOps.exif_transpose() — fixes portrait photos stored with EXIF
  rotation metadata (most phone photos); without this Tesseract reads a
  rotated image and produces garbage
- Upscale images < 600px wide for better character recognition
- Raise binarization threshold 140→160 for faint thermal-print receipts
- Try PSM 6 (single text block) before PSM 4, PSM 11 as fallbacks;
  PSM 6 is better suited to single-column receipt layout

Amount extraction (expenses_agent.py):
- Add Pass 2 bottom-of-receipt line scan when labeled Total: regex fails;
  reads lines bottom-to-top in the last 50% of text, skipping change/tip
  lines — handles 'T0TAL' OCR misread and amount-on-next-line layout
- Add _SKIP_LINE_RE and _ANY_DOLLAR_RE module-level patterns
- 8 new tests covering garbled total, change-skip, USD suffix, etc.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Carlos Garcia
2026-05-20 23:33:38 -04:00
parent f1a8add84b
commit 1536d83376
3 changed files with 100 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
agent_service/agents/expenses_agent.py
View File

@@ -12,7 +12,7 @@ from ..tools.expenses_tools import ExpensesTools
# Receipt OCR helpers — regex-based, deterministic extraction
# ---------------------------------------------------------------------------
# Matches the final-total line on a receipt.
# Matches an explicitly labeled total line.
# Handles "Total: $22.46", "GRAND TOTAL 22.46", "Amount Due: 22.46", etc.
_TOTAL_RE = re.compile(
r'(?:grand\s*total|total\s*due|amount\s*due|balance\s*due|'
@@ -21,22 +21,60 @@ _TOTAL_RE = re.compile(
re.IGNORECASE,
)
# Lines printed AFTER the total (change given, tip, etc.) — skip these
# when doing the bottom-of-receipt scan so we don't mistake them for the total.
_SKIP_LINE_RE = re.compile(
r'\b(?:change|cash\s*(?:paid|tendered)?|tip|gratuity|approved|'
r'auth(?:orized)?|visa|mastercard|amex|discover)\b',
re.IGNORECASE,
)
# Any standalone dollar-like amount (optional $, up to 6 digits, 2 decimals)
_ANY_DOLLAR_RE = re.compile(r'(?<!\d)\$?\s*([\d,]{1,6}\.\d{2})(?!\d)')
_DATE_ISO_RE = re.compile(r'\b(\d{4})[-/](\d{2})[-/](\d{2})\b') # YYYY-MM-DD or YYYY/MM/DD
_DATE_US_RE = re.compile(r'\b(\d{1,2})[/\-](\d{1,2})[/\-](\d{4})\b') # M/D/YYYY
_DATE_US_SHORT_RE = re.compile(r'\b(\d{1,2})[/\-](\d{1,2})[/\-](\d{2})\b') # M/D/YY
def _extract_amount_from_text(text: str) -> float:
"""Return the final total from OCR receipt text, or 0.0 if not found."""
"""Return the final total from OCR receipt text, or 0.0 if not found.
Pass 1 — labeled total: 'Total:', 'Grand Total:', 'Amount Due:', etc.
Pass 2 — bottom scan: reads lines from the bottom of the last 50% of text,
skipping change/cash/tip lines. Handles cases where Tesseract
garbled 'TOTAL' (e.g. 'T0TAL') or placed the amount on its own
line below the label.
"""
if not text:
return 0.0
# Pass 1: explicit label match
matches = list(_TOTAL_RE.finditer(text))
if matches:
raw = matches[-1].group(1).replace(',', '') # last match = grand total
raw = matches[-1].group(1).replace(',', '')
try:
return float(raw)
val = float(raw)
if val > 0:
return val
except ValueError:
pass
# Pass 2: bottom-of-receipt line scan
# Only search the bottom half so item prices (middle section) are excluded
bottom = text[max(0, int(len(text) * 0.5)):]
for line in reversed(bottom.splitlines()):
if _SKIP_LINE_RE.search(line):
continue
m = _ANY_DOLLAR_RE.search(line)
if m:
try:
val = float(m.group(1).replace(',', ''))
if val > 0:
return val
except ValueError:
pass
return 0.0

50
agent_service/tools/receipt_parser.py
View File

@@ -85,34 +85,58 @@ def _ocr_image(data: bytes, filename: str) -> str:
def _ocr_image_tesseract(data: bytes, filename: str) -> str:
"""Tesseract-based OCR pipeline (fallback)."""
"""Tesseract-based OCR pipeline with phone-photo preprocessing."""
try:
from PIL import Image, ImageFilter, ImageOps
import pytesseract
img = Image.open(io.BytesIO(data))
# Resize very large images — tesseract is slower and less accurate at
# phone-camera resolution; 1800px wide is plenty for receipt text.
# ── Step 1: EXIF rotation correction ─────────────────────────────────
# Phone photos are stored with EXIF orientation metadata but the pixel
# data is not actually rotated. Without this fix Tesseract reads a
# portrait receipt as a landscape image and produces garbage.
try:
img = ImageOps.exif_transpose(img)
except Exception:
pass # exif_transpose requires Pillow >= 6.0
# ── Step 2: Resize to working width (1800px) ──────────────────────────
max_w = 1800
if img.width > max_w:
scale = max_w / img.width
img = img.resize((max_w, int(img.height * scale)), Image.LANCZOS)
# Upscale very small images — Tesseract accuracy drops below ~600px
elif img.width < 600:
scale = 600 / img.width
img = img.resize((600, int(img.height * scale)), Image.LANCZOS)
# Grayscale + adaptive binarisation + sharpen
# ── Step 3: Grayscale + contrast ─────────────────────────────────────
img = ImageOps.grayscale(img)
img = ImageOps.autocontrast(img)
img = img.point(lambda x: 0 if x < 140 else 255)
# ── Step 4: Sharpen then binarize ─────────────────────────────────────
# Sharpen first so edges are crisp before thresholding.
# Threshold 160 (was 140) — gentler for faint thermal-print receipts
# where light gray text would be wiped out by the stricter threshold.
img = img.filter(ImageFilter.SHARPEN)
img = img.point(lambda x: 0 if x < 160 else 255)
# psm 1 = automatic page segmentation + OSD (handles rotated receipts).
# Fall back to psm 6 if OSD data is missing.
try:
text = pytesseract.image_to_string(img, config='--oem 3 --psm 1').strip()
except Exception:
text = pytesseract.image_to_string(img, config='--oem 3 --psm 6').strip()
# ── Step 5: OCR — try PSM modes best-suited for receipt layout ────────
# PSM 6 = single uniform text block (best for single-column receipts)
# PSM 4 = single column, variable text sizes (wider fallback)
# PSM 11 = sparse text — last resort for badly segmented images
for psm in (6, 4, 11):
try:
text = pytesseract.image_to_string(
img, config=f'--oem 3 --psm {psm}').strip()
if len(text) >= 20:
logger.debug('Tesseract OCR %s: psm=%d %d chars', filename, psm, len(text))
return text
except Exception:
pass
logger.debug('Tesseract OCR %s: %d chars', filename, len(text))
return text
logger.warning('Tesseract OCR %s: all PSM modes returned < 20 chars', filename)
return ''
except ImportError:
logger.warning('pytesseract/Pillow not installed — OCR unavailable for %s', filename)
return f'[Image: {filename} — install pytesseract+Pillow for OCR]'

21
tests/test_expenses_agent.py
View File

@@ -458,6 +458,27 @@ class TestExtractAmount:
def test_comma_in_amount(self):
assert _extract_amount_from_text('Grand Total: $1,234.56') == 1234.56
def test_bottom_scan_garbled_total(self):
# OCR garbled "TOTAL" — bottom-scan fallback should find the amount
text = 'Burger 5.99\nFries 2.50\nT0TAL 8.49'
assert _extract_amount_from_text(text) == 8.49
def test_bottom_scan_skips_change(self):
# Should return the total (8.49), not the change (1.51)
text = 'TOTAL 8.49\nCash 10.00\nChange 1.51'
assert _extract_amount_from_text(text) == 8.49
def test_bottom_scan_amount_on_own_line(self):
# Amount printed on a separate line below the label
text = 'Items 5.00\nTax 0.50\nTotal\n5.50'
assert _extract_amount_from_text(text) == 5.50
def test_amount_due_with_usd_suffix(self):
# PDF text may include "USD" after the number — regex should still work
# via the bottom scan since the labeled-total regex won't match "USD"
text = 'Total Charged: $198.40 USD'
assert _extract_amount_from_text(text) == 198.40
class TestExtractDate:
def test_iso_format(self):