BillAja / src /models /ai_engine.py
Kndeh
fix: bill calc
be7986f
Raw
History Blame Contribute Delete
41.7 kB
import cv2
import numpy as np
from PIL import Image
import torch
from transformers import DonutProcessor, VisionEncoderDecoderModel
import re
import pandas as pd
import json
# Try importing pytesseract (optional dependency)
try:
import pytesseract
import shutil
if not shutil.which("tesseract"):
import os
default_path = r"C:\Program Files\Tesseract-OCR\tesseract.exe"
if os.path.exists(default_path):
pytesseract.pytesseract.tesseract_cmd = default_path
TESSERACT_AVAILABLE = True
except ImportError:
TESSERACT_AVAILABLE = False
# ===============================
# MODEL LOADING
# ===============================
def load_donut_model(model_path="./donut-mega-finetuned-final-v6", hf_fallback="naver-clova-ix/donut-base-finetuned-cord-v2"):
"""Load Donut model from local path; fall back to HuggingFace Hub if not found."""
import os
# Determine which path to use
if model_path and os.path.isdir(model_path):
load_path = model_path
elif hf_fallback:
load_path = hf_fallback
else:
return None, None, "cpu", False
try:
processor = DonutProcessor.from_pretrained(load_path)
model = VisionEncoderDecoderModel.from_pretrained(load_path)
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
return processor, model, device, True
except Exception as e:
print(f"Failed to load model from {load_path}: {e}")
return None, None, "cpu", False
# ===============================
# PERSPECTIVE CORRECTION
# ===============================
def order_points(pts):
rect = np.zeros((4, 2), dtype=np.float32)
s = pts.sum(axis=1)
rect[0] = pts[np.argmin(s)]
rect[2] = pts[np.argmax(s)]
diff = np.diff(pts, axis=1)
rect[1] = pts[np.argmin(diff)]
rect[3] = pts[np.argmax(diff)]
return rect
def auto_crop_bright_region(img_cv, pad=20):
"""Fallback crop: finds the largest bright (paper-like) region via Otsu."""
try:
gray = cv2.cvtColor(img_cv, cv2.COLOR_BGR2GRAY)
_, mask = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
# Clean noise
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (15, 15))
mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel)
mask = cv2.morphologyEx(mask, cv2.MORPH_OPEN, kernel)
contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if not contours:
return None
h, w = img_cv.shape[:2]
cnt = max(contours, key=cv2.contourArea)
area = cv2.contourArea(cnt)
# Reject if region is too small or covers the whole frame
if area < 0.05 * h * w or area > 0.95 * h * w:
return None
x, y, cw, ch = cv2.boundingRect(cnt)
x0 = max(0, x - pad); y0 = max(0, y - pad)
x1 = min(w, x + cw + pad); y1 = min(h, y + ch + pad)
return img_cv[y0:y1, x0:x1]
except Exception:
return None
def perspective_correction(img_cv):
try:
gray = cv2.cvtColor(img_cv, cv2.COLOR_BGR2GRAY)
blur = cv2.GaussianBlur(gray, (5, 5), 0)
edges = cv2.Canny(blur, 50, 150)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
edges = cv2.dilate(edges, kernel, iterations=2)
contours, _ = cv2.findContours(edges, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if not contours:
return None
contours = sorted(contours, key=cv2.contourArea, reverse=True)
for contour in contours[:5]:
peri = cv2.arcLength(contour, True)
approx = cv2.approxPolyDP(contour, 0.02 * peri, True)
if len(approx) == 4:
pts = approx.reshape(4, 2).astype(np.float32)
rect = order_points(pts)
wA = np.linalg.norm(rect[2] - rect[3])
wB = np.linalg.norm(rect[1] - rect[0])
maxW = int(max(wA, wB))
hA = np.linalg.norm(rect[1] - rect[2])
hB = np.linalg.norm(rect[0] - rect[3])
maxH = int(max(hA, hB))
dst = np.array([[0,0],[maxW-1,0],[maxW-1,maxH-1],[0,maxH-1]], dtype=np.float32)
M = cv2.getPerspectiveTransform(rect, dst)
return cv2.warpPerspective(img_cv, M, (maxW, maxH))
return None
except:
return None
# ===============================
# IMAGE PREPROCESSING — Parameterized + Advanced
# ===============================
def preprocess_receipt(img_array, blur_type="Gaussian", blur_kernel=5, blur_sigma=0,
thresh_block=11, thresh_c=2, enable_bilateral=False,
enable_denoise=False, enable_morph=False, enable_sharpen=False,
enable_clahe=False, enable_perspective=False):
if img_array is None:
return [], None
steps = []
img_cv = cv2.cvtColor(img_array, cv2.COLOR_RGB2BGR)
steps.append((img_array, "1. Original"))
if enable_perspective:
corrected = perspective_correction(img_cv)
if corrected is not None:
img_cv = corrected
steps.append((cv2.cvtColor(img_cv, cv2.COLOR_BGR2RGB), "Perspective Corrected"))
gray = cv2.cvtColor(img_cv, cv2.COLOR_BGR2GRAY)
steps.append((cv2.cvtColor(gray, cv2.COLOR_GRAY2RGB), "2. Grayscale"))
if enable_clahe:
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
gray = clahe.apply(gray)
steps.append((cv2.cvtColor(gray, cv2.COLOR_GRAY2RGB), "CLAHE"))
if enable_denoise:
gray = cv2.fastNlMeansDenoising(gray, None, h=10, templateWindowSize=7, searchWindowSize=21)
steps.append((cv2.cvtColor(gray, cv2.COLOR_GRAY2RGB), "Denoised"))
if enable_bilateral:
gray = cv2.bilateralFilter(gray, 9, 75, 75)
steps.append((cv2.cvtColor(gray, cv2.COLOR_GRAY2RGB), "Bilateral Filter"))
k = max(1, int(blur_kernel))
if k % 2 == 0: k += 1
if blur_type == "Median":
blurred = cv2.medianBlur(gray, k)
elif blur_type == "Box":
blurred = cv2.blur(gray, (k, k))
else:
blurred = cv2.GaussianBlur(gray, (k, k), blur_sigma)
steps.append((cv2.cvtColor(blurred, cv2.COLOR_GRAY2RGB), f"Blur ({blur_type} k={k})"))
if enable_sharpen:
kernel_s = np.array([[-1,-1,-1],[-1,9,-1],[-1,-1,-1]])
blurred = cv2.filter2D(blurred, -1, kernel_s)
steps.append((cv2.cvtColor(blurred, cv2.COLOR_GRAY2RGB), "Sharpened"))
block = max(3, int(thresh_block))
if block % 2 == 0: block += 1
thresh = cv2.adaptiveThreshold(blurred, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY, block, int(thresh_c))
steps.append((cv2.cvtColor(thresh, cv2.COLOR_GRAY2RGB), f"Threshold (b={block}, C={int(thresh_c)})"))
if enable_morph:
km = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))
thresh = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, km)
thresh = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, km)
steps.append((cv2.cvtColor(thresh, cv2.COLOR_GRAY2RGB), "Morphological Clean"))
return steps, thresh
# ===============================
# PRICE UTILITIES
# ===============================
def clean_price(val):
"""Parse price string to float. Handles Indonesian format where dot = thousands separator."""
if not val: return 0.0
if isinstance(val, (int, float)): return float(val)
val_str = str(val).strip()
# Remove currency symbols and whitespace
val_str = re.sub(r'[$€£¥₹Rp\s]', '', val_str)
# Remove non-digit prefix
val_str = re.sub(r'^[^\d]+', '', val_str)
# Handle both dot and comma present (e.g., "1.234,56" or "1,234.56")
if '.' in val_str and ',' in val_str:
if val_str.rfind(',') > val_str.rfind('.'):
# European/Indonesian: 1.234,56 -> 1234.56
val_str = val_str.replace('.', '').replace(',', '.')
else:
# US format: 1,234.56 -> 1234.56
val_str = val_str.replace(',', '')
elif ',' in val_str:
parts = val_str.split(',')
if len(parts) == 2 and len(parts[1]) <= 2:
# Comma is decimal separator: "25,00" -> 25.00
val_str = val_str.replace(',', '.')
else:
# Comma is thousands separator: "25,000" -> 25000
val_str = val_str.replace(',', '')
elif '.' in val_str:
parts = val_str.split('.')
if len(parts) == 2 and len(parts[1]) <= 2:
# Dot is decimal separator: "25.50" -> 25.50 (keep as is)
pass
else:
# Dot is thousands separator: "25.000" -> 25000
val_str = val_str.replace('.', '')
match = re.search(r'[\d]+\.?[\d]*', val_str)
if match:
try: return float(match.group())
except: return 0.0
return 0.0
def is_valid_price_string(val):
if val is None: return False
val_str = str(val).strip()
if not re.search(r'\d', val_str): return False
digits = len(re.findall(r'\d', val_str))
total_alnum = len(re.findall(r'[a-zA-Z\d]', val_str))
if total_alnum > 0 and digits / total_alnum < 0.3: return False
return True
def clean_item_name(nm):
if isinstance(nm, list):
cleaned_elements = []
for x in nm:
x_str = re.sub(r'<.*?>', '', str(x)).strip()
if x_str:
cleaned_elements.append(x_str)
nm = " ".join(cleaned_elements)
nm_str = str(nm).strip()
# Remove leading non-word characters like colons, slashes, periods, spaces
nm_str = re.sub(r'^[^\w]+', '', nm_str).strip()
return nm_str
def is_valid_item(nm, price_str, cnt_str=None):
if not nm: return False
nm_str = clean_item_name(nm)
if len(nm_str) < 2: return False
price_raw = str(price_str).strip()
cnt_raw = str(cnt_str).strip() if cnt_str is not None else ""
for val_check in [price_raw, cnt_raw]:
if not val_check: continue
if re.search(r'\d{1,2}:\d{2}', val_check): return False
if re.search(r'\d{1,4}[/\-\.]\d{1,2}[/\-\.]\d{2,4}', val_check): return False
if len(re.findall(r'\d', val_check)) > 8 and ('/' in val_check or '-' in val_check or len(val_check) > 10):
return False
if not is_valid_price_string(price_str): return False
price = clean_price(price_str)
if price <= 0: return False
skip_patterns = [
# English terms
r'(?i)^invoice', r'(?i)^date\s*(of|:)', r'(?i)^seller\s*:?', r'(?i)^client\s*:?',
r'(?i)^buyer\s*:?', r'(?i)^customer\s*:?', r'(?i)^tax\s*id', r'(?i)^iban\s*:?',
r'(?i)^dpo\s', r'(?i)^items?\s*$', r'(?i)^total\s*$', r'(?i)^sub\s*total',
r'(?i)^summary', r'(?i)^vat\s', r'(?i)^no\.\s*$', r'(?i)^description\s*$',
r'(?i)^worth\s*$', r'(?i)^ibay\s*:?', r'(?i)^qty\s*$', r'(?i)^quantity\s*$',
r'(?i)^unit\s*price', r'(?i)^net\s*(worth|price)', r'(?i)^gross\s*(worth|price)',
r'(?i)^amount\s*$', r'(?i)^payment', r'(?i)^change\s*$', r'(?i)^cash\s*$',
r'(?i)^credit\s*card', r'(?i)^thank\s*you', r'(?i)^receipt',
r'(?i)^bill\s*(no|number)', r'(?i)^order\s*(no|number|id)',
r'(?i)^table\s*(no|number)', r'(?i)^server\s*:?', r'(?i)^cashier\s*:?',
r'(?i)^discount', r'(?i)^bill\s*discount',
# Indonesian terms
r'(?i)^kasir', r'(?i)^pelayan', r'(?i)^meja', r'(?i)^nomor', r'(?i)^no\b',
r'(?i)^tanggal', r'(?i)^jam\b', r'(?i)^telp', r'(?i)^phone', r'(?i)^telepon',
r'(?i)^alamat', r'(?i)^ruko', r'(?i)^mall', r'(?i)^lantai', r'(?i)^floor',
r'(?i)^kota', r'(?i)^jalan', r'(?i)^jl\b', r'(?i)^kembali', r'(?i)^kembalian',
r'(?i)^tunai', r'(?i)^debit', r'(?i)^kredit', r'(?i)^lunas', r'(?i)^pajak',
r'(?i)^ppn', r'(?i)^diskon', r'(?i)^potongan', r'(?i)^promo', r'(?i)^voucher',
r'(?i)^qris', r'(?i)^bca', r'(?i)^mandiri', r'(?i)^bri', r'(?i)^bni',
r'(?i)^ovo', r'(?i)^gopay', r'(?i)^dana', r'(?i)^linkaja', r'(?i)^merchant',
]
for p in skip_patterns:
if re.search(p, nm_str): return False
if cnt_str is not None:
cnt_val = clean_price(cnt_str)
cnt_clean = re.sub(r'[^\w\.\,\/\-]', '', cnt_raw)
if len(cnt_clean) > 5 and ('/' in cnt_clean or '-' in cnt_clean):
return False
if cnt_val > 1000:
return False
price_raw_clean = str(price_str).strip()
# Strip known currency prefix (Rp, $, etc.) then reject if ANY letter remains in price
price_no_currency = re.sub(r'^[Rr][Pp]\.?\s*|^[$€£¥]\s*', '', price_raw_clean)
if re.search(r'[a-zA-Z]', price_no_currency):
return False
# Reject only obvious transaction codes / serial IDs:
# must be all-caps, contain digits, AND contain a separator like / or - or be >= 8 digits
if re.match(r'^[A-Z0-9\s\-:/\.]{5,}$', nm_str) and not re.search(r'[a-z]', nm_str):
digit_count = len(re.findall(r'\d', nm_str))
has_separator = bool(re.search(r'[/\-:]', nm_str))
# Only reject if it looks like a code: has digits AND a separator, OR is >50% digits
word_count = len(nm_str.split())
if digit_count > 0 and has_separator:
return False
if word_count <= 2 and digit_count > 0 and digit_count >= len(nm_str.replace(' ', '')) * 0.4:
return False
return True
# ===============================
# SMART FIELD MAPPING FIX
# ===============================
def smart_fix_summary(items_sum, subtotal, tax, service, discount, total):
"""Fix commonly misassigned CORD sub_total/total fields using heuristics."""
vals = {"subtotal": subtotal, "tax": tax, "service": service, "total": total}
non_zero = {k: v for k, v in vals.items() if v > 0}
if not non_zero:
return subtotal, tax, service, discount, total
# Rule 1: largest value is most likely the total
sorted_v = sorted(non_zero.items(), key=lambda x: x[1], reverse=True)
top_key, top_val = sorted_v[0]
if top_key != "total" and top_val > total:
old_total = total
total = top_val
if top_key == "service": service = old_total
elif top_key == "tax": tax = old_total
elif top_key == "subtotal": subtotal = old_total
# Rule 2: service should be small; if > items_sum it's wrong
if service > 0 and items_sum > 0 and service > items_sum * 0.5:
if abs(service - total) < 2:
service = 0.0
elif service > total:
service, total = 0.0, service
# Rule 3: if total == tax and there's a bigger value elsewhere, fix it
if total > 0 and total == tax:
candidates = [v for k, v in non_zero.items() if k not in ("tax", "total") and v > total]
if candidates:
total = max(candidates)
service = 0.0
# Rule 4: derive subtotal from items if missing
if subtotal == 0 and items_sum > 0:
subtotal = items_sum
# Rule 5: if total < items_sum, the total is clearly wrong — recompute from items + extras
if items_sum > 0 and 0 < total < items_sum * 0.8:
computed = items_sum + tax + service - discount
if computed > total:
total = computed
return subtotal, tax, service, discount, total
# ===============================
# DONUT PARSING
# ===============================
def parse_cord_to_schema(cord_json):
if not isinstance(cord_json, dict):
cord_json = {}
items = []
menu = cord_json.get("menu", [])
if isinstance(menu, dict): menu = [menu]
tax_from_items = 0.0
service_from_items = 0.0
discount_from_items = 0.0
for item in menu:
if not isinstance(item, dict):
continue
for entry in [item]:
if not isinstance(entry, dict):
continue
nm = entry.get("nm", None)
cnt_field = entry.get("cnt", "1")
price_field = entry.get("price", "0")
# Recover nm from malformed structure where cnt contains a nested dict with nm
if nm is None and isinstance(cnt_field, dict):
nm = cnt_field.get("nm", None)
cnt_field = "1"
# Skip items with no usable name
if not nm or str(nm).strip() in ("", "Unknown"):
continue
# Flatten cnt: if it's a dict, try to get a numeric string from it
if isinstance(cnt_field, dict):
cnt_field = next(
(str(v) for v in cnt_field.values() if v and str(v).strip().replace('.','').replace(',','').isdigit()),
"1"
)
cnt_raw = str(cnt_field).strip() if cnt_field else "1"
# Flatten price: if it's a dict, the price field got corrupted — skip this entry
if isinstance(price_field, dict):
continue
price_raw = str(price_field).strip() if price_field else "0"
if not is_valid_item(nm, price_raw, cnt_raw):
if nm:
nm_lower = clean_item_name(nm).lower()
if any(k in nm_lower for k in ("discount", "diskon", "potongan", "promo", "voucher")):
discount_from_items += clean_price(price_raw)
elif any(k in nm_lower for k in ("tax", "vat", "pajak", "ppn")):
tax_from_items += clean_price(price_raw)
elif any(k in nm_lower for k in ("service", "servis", "sc", "charge")):
service_from_items += clean_price(price_raw)
continue
price = clean_price(price_raw)
cnt = clean_price(cnt_raw)
if cnt <= 0 or cnt > 1000: cnt = 1
items.append({"item_name": clean_item_name(nm), "item_quantity": cnt, "item_price": price})
sub_total_node = cord_json.get("sub_total", {})
if isinstance(sub_total_node, list):
sub_total_node = sub_total_node[0] if sub_total_node else {}
if not isinstance(sub_total_node, dict):
sub_total_node = {}
subtotal = clean_price(sub_total_node.get("subtotal_price", "0"))
tax_amount = clean_price(sub_total_node.get("tax_price", "0"))
service_charge = clean_price(sub_total_node.get("service_price", "0"))
discount_val = clean_price(sub_total_node.get("discount_price", "0"))
if tax_amount == 0.0 and tax_from_items > 0.0:
tax_amount = tax_from_items
if service_charge == 0.0 and service_from_items > 0.0:
service_charge = service_from_items
if discount_val == 0.0 and discount_from_items > 0.0:
discount_val = discount_from_items
total_node = cord_json.get("total", {})
if isinstance(total_node, list):
total_node = total_node[0] if total_node else {}
if not isinstance(total_node, dict):
total_node = {}
total_amount = clean_price(total_node.get("total_price", "0"))
items_sum = sum(i["item_price"] for i in items)
subtotal, tax_amount, service_charge, discount_val, total_amount = smart_fix_summary(
items_sum, subtotal, tax_amount, service_charge, discount_val, total_amount
)
return {
"items": items, "subtotal": subtotal, "tax_amount": tax_amount,
"service_charge": service_charge,
"discount_details": {"type": "fixed" if discount_val > 0 else "none", "value": discount_val},
"total_amount": total_amount
}
# ===============================
# PHASE 2 - SENSITIVE DATA FILTER
# ===============================
SENSITIVE_FIELDS = {"cashprice", "changeprice", "creditcardprice", "emoneyprice",
"sub_nm", "sub_price", "sub_cnt", "sub_etc"}
def filter_sensitive_fields(cord_json):
filtered = {}
if "menu" in cord_json:
menu = cord_json["menu"]
if isinstance(menu, dict):
menu = [menu]
clean_menu = []
for item in menu:
if isinstance(item, dict):
clean_item = {k: v for k, v in item.items()
if k in ("nm", "name", "price", "cnt", "unitprice", "itemsubtotal")}
if clean_item:
clean_menu.append(clean_item)
filtered["menu"] = clean_menu
if "sub_total" in cord_json and isinstance(cord_json["sub_total"], dict):
filtered["sub_total"] = {"subtotal_price": cord_json["sub_total"].get("subtotal_price", "")}
if "total" in cord_json and isinstance(cord_json["total"], dict):
filtered["total"] = {"total_price": cord_json["total"].get("total_price", "")}
return filtered
# ===============================
# DONUT OCR
# ===============================
def run_donut_ocr(img_array, processor, model, device, model_loaded,
blur_type="Gaussian", blur_kernel=5, blur_sigma=0,
thresh_block=11, thresh_c=2, use_preprocessed=False,
**filter_flags):
gallery, thresh = preprocess_receipt(img_array, blur_type, blur_kernel, blur_sigma,
thresh_block, thresh_c, **filter_flags)
if not model_loaded:
return gallery, {"error": "Donut model not loaded"}, None
if use_preprocessed and thresh is not None:
# Feed the thresholded (binary) preprocessing output as RGB
rgb_for_donut = cv2.cvtColor(thresh, cv2.COLOR_GRAY2RGB) if len(thresh.shape) == 2 else thresh
print(f"[Donut] Using preprocessed image, shape={rgb_for_donut.shape}")
else:
# Use RGB image, but apply perspective fix if requested
if filter_flags.get("enable_perspective", False):
try:
bgr = cv2.cvtColor(img_array, cv2.COLOR_RGB2BGR)
cropped = perspective_correction(bgr)
if cropped is None:
cropped = auto_crop_bright_region(bgr)
if cropped is not None:
rgb_for_donut = cv2.cvtColor(cropped, cv2.COLOR_BGR2RGB)
print(f"[Donut] Cropped {bgr.shape[:2]} -> {cropped.shape[:2]}")
else:
rgb_for_donut = img_array
except Exception as e:
rgb_for_donut = img_array
print(f"[Donut] crop failed: {e}")
else:
rgb_for_donut = img_array
pil_img = Image.fromarray(rgb_for_donut).convert("RGB")
# Let the processor handle resizing to the model's expected input size
pixel_values = processor(pil_img, return_tensors="pt").pixel_values.to(device)
task_prompt = "<s_cord-v2>"
decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids.to(device)
outputs = model.generate(
pixel_values, decoder_input_ids=decoder_input_ids,
max_new_tokens=model.decoder.config.max_position_embeddings,
pad_token_id=processor.tokenizer.pad_token_id,
eos_token_id=processor.tokenizer.eos_token_id,
use_cache=True, bad_words_ids=[[processor.tokenizer.unk_token_id]],
return_dict_in_generate=True,
)
sequence = processor.batch_decode(outputs.sequences)[0]
sequence = sequence.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
raw_seq = sequence
sequence_stripped = re.sub(r"<.*?>", "", sequence, count=1).strip()
def fallback_regex_parse(sequence):
"""
Fallback parser that uses regex to extract menu items from malformed XML sequences.
This works when processor.token2json() fails due to unclosed tags or syntax mismatch.
"""
menu_match = re.search(r'<s_menu>(.*?)</s_menu>', sequence)
if menu_match:
menu_content = menu_match.group(1)
else:
menu_split = sequence.split('<s_menu>')
if len(menu_split) > 1:
menu_content = menu_split[1]
else:
menu_content = sequence
def extract_field(tag, src):
m = re.search(rf'<s_{tag}_price>(.*?)</s_{tag}_price>', src)
if not m:
m = re.search(rf'<s_{tag}>(.*?)</s_{tag}>', src)
return m.group(1).strip() if m else ""
subtotal_str = extract_field("subtotal", sequence)
tax_str = extract_field("tax", sequence)
service_str = extract_field("service", sequence)
discount_str = extract_field("discount", sequence)
total_str = extract_field("total", sequence)
parts = re.split(r'<sep/>|<s_nm>', menu_content)
menu_items = []
for part in parts:
part = part.strip()
if not part: continue
nm_m = re.match(r'^([^<]+)', part)
nm_val = nm_m.group(1).strip() if nm_m else ""
if not nm_val:
nm_m2 = re.search(r'^(.*?)<\/s_nm>', part)
if nm_m2:
nm_val = nm_m2.group(1).strip()
nm_val = re.sub(r'<.*?>', '', nm_val).strip()
# Price: try <s_price>...</s_price>, then <s_price>...<, then trailing digits before </s_price>
price_val = ""
price_m = re.search(r'<s_price>(.*?)(?:</s_price>|<s_[a-z])', part)
if price_m:
price_val = re.sub(r'<.*?>', '', price_m.group(1)).strip()
if not price_val:
# Match digits/commas/dots just before </s_price> even without opening tag
price_m2 = re.search(r'(\d[\d\.,]*)\s*</s_price>', part)
if price_m2: price_val = price_m2.group(1).strip()
if not price_val:
price_m3 = re.search(r'<s_price>([^<]+)', part)
if price_m3: price_val = price_m3.group(1).strip()
# Cnt: try <s_cnt>...</s_cnt>, then <s_cnt>...<
cnt_val = ""
cnt_m = re.search(r'<s_cnt>(.*?)(?:</s_cnt>|<s_[a-z])', part)
if cnt_m:
cnt_val = re.sub(r'<.*?>', '', cnt_m.group(1)).strip()
if not cnt_val:
cnt_m2 = re.search(r'<s_cnt>([^<]+)', part)
if cnt_m2: cnt_val = cnt_m2.group(1).strip()
if nm_val or price_val:
menu_items.append({
"nm": nm_val,
"price": price_val,
"cnt": cnt_val
})
cord_json = {
"menu": menu_items,
"sub_total": {
"subtotal_price": subtotal_str,
"tax_price": tax_str,
"service_price": service_str,
"discount_price": discount_str
},
"total": {
"total_price": total_str
}
}
return cord_json
# ===============================
# DONUT OCR
# ===============================
def run_donut_ocr(img_array, processor, model, device, model_loaded,
blur_type="Gaussian", blur_kernel=5, blur_sigma=0,
thresh_block=11, thresh_c=2, use_preprocessed=False,
**filter_flags):
gallery, thresh = preprocess_receipt(img_array, blur_type, blur_kernel, blur_sigma,
thresh_block, thresh_c, **filter_flags)
if not model_loaded:
return gallery, {"error": "Donut model not loaded"}, None
if use_preprocessed and thresh is not None:
rgb_for_donut = cv2.cvtColor(thresh, cv2.COLOR_GRAY2RGB) if len(thresh.shape) == 2 else thresh
print(f"[Donut] Using preprocessed image, shape={rgb_for_donut.shape}")
else:
if filter_flags.get("enable_perspective", False):
try:
bgr = cv2.cvtColor(img_array, cv2.COLOR_RGB2BGR)
cropped = perspective_correction(bgr)
if cropped is None:
cropped = auto_crop_bright_region(bgr)
if cropped is not None:
rgb_for_donut = cv2.cvtColor(cropped, cv2.COLOR_BGR2RGB)
print(f"[Donut] Cropped {bgr.shape[:2]} -> {cropped.shape[:2]}")
else:
rgb_for_donut = img_array
except Exception as e:
rgb_for_donut = img_array
print(f"[Donut] crop failed: {e}")
else:
rgb_for_donut = img_array
pil_img = Image.fromarray(rgb_for_donut).convert("RGB")
pixel_values = processor(pil_img, return_tensors="pt").pixel_values.to(device)
task_prompt = "<s_cord-v2>"
decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids.to(device)
outputs = model.generate(
pixel_values, decoder_input_ids=decoder_input_ids,
max_new_tokens=model.decoder.config.max_position_embeddings,
pad_token_id=processor.tokenizer.pad_token_id,
eos_token_id=processor.tokenizer.eos_token_id,
use_cache=True, bad_words_ids=[[processor.tokenizer.unk_token_id]],
return_dict_in_generate=True,
)
sequence = processor.batch_decode(outputs.sequences)[0]
sequence = sequence.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
raw_seq = sequence
sequence_stripped = re.sub(r"<.*?>", "", sequence, count=1).strip()
raw_json = processor.token2json(sequence_stripped)
if isinstance(raw_json, list):
raw_json = {"menu": raw_json}
# Normalize menu to list if it's a dict
if isinstance(raw_json, dict):
menu_node = raw_json.get("menu")
if isinstance(menu_node, dict):
raw_json["menu"] = [menu_node]
# Always also run regex fallback and merge any items it finds that token2json missed
# Only merge items that have BOTH a valid name AND a numeric-looking price (prevent junk)
regex_json = fallback_regex_parse(sequence)
token_names = {clean_item_name(e.get("nm", "")).lower()
for e in (raw_json.get("menu", []) if isinstance(raw_json, dict) else [])
if isinstance(e, dict) and e.get("nm")}
for extra in regex_json.get("menu", []):
if not isinstance(extra, dict): continue
extra_nm = clean_item_name(extra.get("nm", ""))
extra_price = extra.get("price", "")
# Guard: only merge if price string is actually numeric (has digits, no stray letters)
if not extra_nm or not is_valid_price_string(extra_price):
continue
price_no_curr = re.sub(r'^[Rr][Pp]\.?\s*|^[$€£¥]\s*', '', str(extra_price).strip())
if re.search(r'[a-zA-Z]', price_no_curr):
continue
if extra_nm.lower() not in token_names:
if not isinstance(raw_json, dict): raw_json = {}
raw_json.setdefault("menu", []).append(extra)
token_names.add(extra_nm.lower())
# Fill missing sub_total / total from regex if token2json missed them
if isinstance(raw_json, dict):
for section in ("sub_total", "total"):
if not raw_json.get(section) and regex_json.get(section):
raw_json[section] = regex_json[section]
# Primary fallback: use full regex result if token2json produced nothing
is_fallback = False
if not isinstance(raw_json, dict) or not raw_json.get("menu"):
try:
candidate = re.sub(r"^<[^>]+>", "", sequence.strip()).rstrip("</s>").strip()
raw_json = json.loads(candidate)
if isinstance(raw_json, list):
raw_json = {"menu": raw_json}
print("[Donut] Parsed output as raw JSON (fine-tuned model format)")
except Exception:
raw_json = regex_json
is_fallback = True
print("[Donut] Parsed output using robust regex fallback parser")
parsed = parse_cord_to_schema(raw_json)
gallery = list(gallery) + [(rgb_for_donut, "→ Donut Input (cropped)")]
return gallery, {
"ocr_engine": "Donut (Regex Fallback)" if is_fallback else "Donut",
"raw_sequence": raw_seq,
"raw_cord": filter_sensitive_fields(raw_json),
"parsed": parsed
}, parsed
# ===============================
# TESSERACT OCR
# ===============================
def parse_tesseract_text(raw_text):
lines = raw_text.strip().split('\n')
items = []
tax, service, discount, total, subtotal = 0.0, 0.0, 0.0, 0.0, 0.0
summary_pats = {
'tax': r'(?i)(?:tax|vat|pajak|ppn)\s*[:\s]*[\$€£¥₹Rp\s]*([0-9][0-9.,]*)',
'service': r'(?i)(?:service\s*charge|servis|sc)\s*[:\s]*[\$€£¥₹Rp\s]*([0-9][0-9.,]*)',
'discount': r'(?i)(?:discount|diskon|potongan)\s*[:\s]*[\$€£¥₹Rp\s]*([0-9][0-9.,]*)',
'total': r'(?i)(?:(?:grand\s*)?total)\s*[:\s]*[\$€£¥₹Rp\s]*([0-9][0-9.,]*)',
'subtotal': r'(?i)(?:sub\s*total)\s*[:\s]*[\$€£¥₹Rp\s]*([0-9][0-9.,]*)',
}
skip_pats = [
r'(?i)^\s*invoice', r'(?i)^\s*date\s*(of|:)', r'(?i)^\s*seller', r'(?i)^\s*client',
r'(?i)^\s*buyer', r'(?i)^\s*customer', r'(?i)^\s*tax\s*id', r'(?i)^\s*iban',
r'(?i)^\s*phone', r'(?i)^\s*tel\s*:?', r'(?i)^\s*address', r'(?i)^\s*thank\s*you',
r'(?i)^\s*receipt', r'(?i)^\s*bill\s*(no|num)', r'(?i)^\s*order\s*(no|num)',
r'(?i)^\s*table\s*(no|num)', r'(?i)^\s*server', r'(?i)^\s*cashier',
r'(?i)^\s*payment', r'(?i)^\s*change', r'(?i)^\s*cash\s*:?',
r'(?i)^\s*credit\s*card', r'(?i)^\s*summary\s*$', r'(?i)^\s*items?\s*$',
r'(?i)^\s*no\.\s+desc', r'(?i)^\s*qty\s+', r'^\s*[-=*]{3,}', r'^\s*$',
]
for line in lines:
line = line.strip()
if not line: continue
is_summary = False
for key, pat in summary_pats.items():
m = re.search(pat, line)
if m:
v = clean_price(m.group(1))
if key == 'tax': tax = v
elif key == 'service': service = v
elif key == 'discount': discount = v
elif key == 'total': total = v
elif key == 'subtotal': subtotal = v
is_summary = True; break
if is_summary: continue
skip = False
for p in skip_pats:
if re.search(p, line): skip = True; break
if skip: continue
numbers = re.findall(r'[\d][0-9.,]*[\d]|[\d]+', line)
if not numbers: continue
price = clean_price(numbers[-1])
if price <= 0: continue
first_num = re.search(r'\s+[\d]', line)
item_name = line[:first_num.start()].strip() if first_num else line.strip()
item_name = re.sub(r'^\d+[\.)\s]+', '', item_name).strip()
if not item_name or len(item_name) < 2: continue
qty = 1.0
qm = re.search(r'(\d+)\s*[xX×]', line)
if qm: qty = float(qm.group(1))
elif len(numbers) >= 3:
pq = clean_price(numbers[0])
if 0 < pq <= 100: qty = pq
items.append({"item_name": item_name, "item_quantity": qty, "item_price": price})
return {
"items": items, "subtotal": subtotal, "tax_amount": tax,
"service_charge": service,
"discount_details": {"type": "fixed" if discount > 0 else "none", "value": discount},
"total_amount": total
}
def run_tesseract_ocr(img_array, blur_type="Gaussian", blur_kernel=5, blur_sigma=0,
thresh_block=11, thresh_c=2, **filter_flags):
gallery, thresh = preprocess_receipt(img_array, blur_type, blur_kernel, blur_sigma,
thresh_block, thresh_c, **filter_flags)
if not TESSERACT_AVAILABLE:
return gallery, {"error": "Tesseract not installed"}, None
if thresh is None:
return gallery, {"error": "No image provided"}, None
try:
raw_text = pytesseract.image_to_string(Image.fromarray(thresh), lang='eng')
parsed = parse_tesseract_text(raw_text)
result_json = {"ocr_engine": "Tesseract", "raw_text": raw_text, "parsed": parsed}
return gallery, result_json, parsed
except Exception as e:
return gallery, {"error": f"Tesseract failed: {e}"}, None
# ===============================
# HELPERS — Populate split bill from parsed data
# ===============================
def parsed_to_split_bill(parsed):
if parsed is None:
return pd.DataFrame(columns=["Item Name","Qty","Price","Assigned To"]), 0, 0, 0, 0
rows = [{"Item Name": i["item_name"], "Qty": i["item_quantity"],
"Price": i["item_price"], "Assigned To": ""} for i in parsed.get("items", [])]
df = pd.DataFrame(rows) if rows else pd.DataFrame(columns=["Item Name","Qty","Price","Assigned To"])
return (df, parsed.get("tax_amount", 0), parsed.get("service_charge", 0),
parsed.get("discount_details", {}).get("value", 0), parsed.get("total_amount", 0))
def add_item(df):
if df is None or df.empty:
df = pd.DataFrame(columns=["Item Name","Qty","Price","Assigned To"])
new = pd.DataFrame([{"Item Name": "", "Qty": 1, "Price": 0, "Assigned To": ""}])
return pd.concat([df, new], ignore_index=True)
def remove_last_item(df):
if df is None or len(df) == 0: return df
return df.iloc[:-1].reset_index(drop=True)
def assign_all_unassigned(df, name):
if df is None or df.empty or not name: return df
df = df.copy()
mask = df["Assigned To"].isna() | (df["Assigned To"].astype(str).str.strip() == "")
df.loc[mask, "Assigned To"] = name.strip()
return df
# ===============================
# OCR ACCURACY COMPARISON
# ===============================
def score_parsed_result(parsed):
"""Score a parsed OCR result 0-50. Higher = more likely accurate."""
if not parsed:
return 0, {}
score = 0
items = parsed.get("items", [])
total = parsed.get("total_amount", 0)
items_sum = sum(i["item_price"] for i in items)
# Up to 20 pts: item count (2 per item, max 10 items)
item_score = min(len(items) * 2, 20)
score += item_score
# 5 pts: total amount was detected
if total > 0:
score += 5
# Up to 15 pts: how close is items_sum to total
sum_ratio = 0.0
if total > 0 and items_sum > 0:
sum_ratio = min(items_sum, total) / max(items_sum, total)
score += int(sum_ratio * 15)
# Up to 10 pts: item name quality (length 3-60, no excessive special chars)
name_quality = 0.0
if items:
ok = sum(
1 for i in items
if 3 <= len(str(i.get("item_name", ""))) <= 60
and len(re.findall(r'[^a-zA-Z0-9\s\-\./,()&]', str(i.get("item_name", "")))) <= 2
)
name_quality = ok / len(items)
score += int(name_quality * 10)
return score, {
"items_found": len(items),
"items_sum": items_sum,
"total_detected": total,
"sum_accuracy_pct": round(sum_ratio * 100, 1),
"name_quality_pct": round(name_quality * 100, 1),
}
def compare_ocr_results(donut_parsed, tess_parsed):
"""Compare Donut and Tesseract results, return recommendation dict."""
donut_score, donut_detail = score_parsed_result(donut_parsed)
tess_score, tess_detail = score_parsed_result(tess_parsed)
gap = abs(donut_score - tess_score)
if donut_score > tess_score:
winner = "donut"
elif tess_score > donut_score:
winner = "tesseract"
else:
winner = "tie"
confidence = "high" if gap >= 10 else ("moderate" if gap >= 4 else "low")
return {
"winner": winner,
"confidence": confidence,
"donut_score": donut_score,
"tess_score": tess_score,
"donut_detail": donut_detail,
"tess_detail": tess_detail,
}
def calculate_split_bill(df, tax, service, discount, total, split_mode="After Tax"):
if df is None or df.empty:
return pd.DataFrame(columns=["Person","Items Total","Tax","Service","Discount","Total to Pay"]), "No items to split."
# Guard against None from empty Streamlit number_input
tax = float(tax or 0)
service = float(service or 0)
discount = float(discount or 0)
total = float(total or 0)
total_items_price = df["Price"].astype(float).sum()
if total_items_price == 0:
return pd.DataFrame(), "Total items price is 0."
person_totals = {}
for _, row in df.iterrows():
price = float(row["Price"])
assigned = str(row.get("Assigned To", "")).strip()
if not assigned: continue
persons = [p.strip() for p in assigned.split(",") if p.strip()]
if not persons: continue
split_price = price / len(persons)
for p in persons:
if p not in person_totals:
person_totals[p] = {"items_cost": 0.0}
person_totals[p]["items_cost"] += split_price
if not person_totals:
return pd.DataFrame(), "No one assigned to any items."
num_people = len(person_totals)
result_list = []
for p, data in person_totals.items():
proportion = data["items_cost"] / total_items_price
if split_mode == "Before Tax (Equal Extras)":
p_tax = tax / num_people
p_svc = service / num_people
p_disc = discount / num_people
else: # After Tax (Proportional)
p_tax = tax * proportion
p_svc = service * proportion
p_disc = discount * proportion
p_total = data["items_cost"] + p_tax + p_svc - p_disc
result_list.append({
"Person": p, "Items Total": round(data["items_cost"], 2),
"Tax": round(p_tax, 2), "Service": round(p_svc, 2),
"Discount": round(p_disc, 2), "Total to Pay": round(p_total, 2)
})
df_res = pd.DataFrame(result_list)
calc_total = sum(r["Total to Pay"] for r in result_list)
summary = f"**Calculated Total:** Rp {calc_total:,.0f} | **Receipt Total:** Rp {total:,.0f}"
if total > 0 and abs(calc_total - total) > 1000:
summary += "\n\n⚠️ *Warning: Calculated total differs significantly from receipt total.*"
return df_res, summary