# extractors/voter.py import re from extractors.base import ExtractResult from utils.text import EPIC_RX, DOB_RX # Common label tokens in english and bengali (and some transliteration) NAME_LABELS = [ r"elector'?s name", r"electors name", r"elector name", r"name", r"নাম", r"নামঃ", r"নামের", r"নামঃ", r"নাম:", r"নাম।" ] FATHER_LABELS = [ r"father'?s name", r"husband'?s name", r"father name", r"পিতা|পিতার নাম", r"স্বামীর নাম", r"স্বামী" ] DOB_LABELS = [ r"dob", r"date of birth", r"birth", r"জন্ম", r"জন্ম তারিখ" ] GENDER_LABELS = [r"male|female|transgender|মহিলা|পুরুষ|পুরুষ"] def _clean_token(s: str) -> str: """Basic cleanup of OCR tokens: trim, remove extra punctuation and common OCR artifacts.""" if not s: return s s = s.strip() # remove weird repeated punctuation and control chars: s = re.sub(r"[\u200b\u200c\u200d]", "", s) s = re.sub(r"[_\uFFFD]+", "", s) s = s.strip(" \t\n\r:.-,") # collapse multiple spaces s = re.sub(r"\s{2,}", " ", s) return s def _looks_like_name(s: str) -> bool: """Rudimentary check — a name is usually 2+ words and not mostly digits.""" if not s: return False s2 = re.sub(r"[^A-Za-z\u0980-\u09FF0-9 ]", "", s) # allow Bengali block too words = [w for w in s2.split() if w.strip()] return len(words) >= 2 and sum(ch.isdigit() for ch in s2) < len(s2) * 0.4 def _join_split_label(lines): """ Fix cases where OCR split a label across lines and produced "'s Name" or similar. If a line looks like " 's Name" or "Name" preceded by something like "Father", join them. """ out = [] i = 0 while i < len(lines): cur = lines[i] # look ahead for patterns like "'s Name" or "Name" that are likely continuation if i + 1 < len(lines): next_line = lines[i+1] # next_line is a short continuation like "'s Name" or "Name" if re.match(r"^[\'\"\u2019sS]{0,3}\s*name[\s\:\-]*$", next_line.strip(), re.I) or re.match(r"^[\'\"\u2019sS]{0,3}\s*নাম[\s\:\-]*$", next_line.strip(), re.I): combined = (cur + " " + next_line).strip() out.append(combined) i += 2 continue out.append(cur) i += 1 return out def extract_voter(text: str) -> ExtractResult: """ Improved voter extractor: - Uses EPIC regex to find card id - Searches neighbor lines for name and father/husband name - Supports English + Bengali labels and fixes common split-label artifacts - Tries several heuristics to guess name (lines around photo/EPIC, label-based extraction) """ if not text: return ExtractResult("voter", {"epic": None, "name": None, "father_or_spouse": None, "dob": None, "gender": None, "address": None}, 0.0) # Normalize newlines and whitespace lines = [l.strip() for l in text.splitlines() if l.strip()] # fix common split cases where label got split into "'s Name" lines = _join_split_label(lines) # Quick lower-case version for label checks llines_lower = [l.lower() for l in lines] epic = None name = None father = None dob = None gender = None address = None # find EPIC first and note its line index epic_idx = None for i, l in enumerate(lines): m = EPIC_RX.search(l) if m: epic = m.group(1) epic_idx = i break # try to find DOB anywhere for l in lines: m = DOB_RX.search(l) if m: dob = _clean_token(m.group(1)) break # try to find gender token for l in lines: if re.search(r"\b(male|female|transgender|male\b|female\b)\b", l, re.I) or re.search(r"মহিলা|পুরুষ|নারী|পুরুষ", l): gender = _clean_token(l) break # label-based extraction (preferred) for i, l in enumerate(lines): low = l.lower() # name labels if any(re.search(lbl, low) for lbl in NAME_LABELS): # strip the label portion and keep the remainder # e.g. "Elector's Name: John Doe" -> "John Doe" try: name_candidate = re.split(r"[:\-]\s*", l, maxsplit=1)[1] except Exception: # fallback: take next line if label line has no immediate name name_candidate = lines[i+1] if i+1 < len(lines) else "" name_candidate = _clean_token(name_candidate) if not name_candidate: # maybe name is in next line if i+1 < len(lines): name_candidate = _clean_token(lines[i+1]) if _looks_like_name(name_candidate): name = name_candidate break # father/husband label based for i, l in enumerate(lines): low = l.lower() if any(re.search(lbl, low) for lbl in FATHER_LABELS): try: candidate = re.split(r"[:\-]\s*", l, maxsplit=1)[1] except Exception: candidate = lines[i+1] if i+1 < len(lines) else "" candidate = _clean_token(candidate) if not candidate and i+1 < len(lines): candidate = _clean_token(lines[i+1]) # sometimes "Father's Name" might be OCRed as "Father's" and next line "'s Name" etc. if candidate: father = candidate break # If the label-based approach didn't find name/father, try positional heuristics relative to EPIC or photo if (not name or not _looks_like_name(name)) and epic_idx is not None: # common layouts: name often appears 1-3 lines above EPIC or just below the title search_range = list(range(max(0, epic_idx-4), epic_idx)) + list(range(epic_idx+1, min(len(lines), epic_idx+4))) for i in search_range: cand = _clean_token(lines[i]) if _looks_like_name(cand): # if name empty, take first plausible candidate; if father empty, subsequent plausible candidate if not name: name = cand elif not father and cand != name: father = cand if name and father: break # Fallback: sometimes first non-label line is name (title area) if not name: for i in range(min(6, len(lines))): l = _clean_token(lines[i]) # skip lines that are clearly labels or single tokens like "ELECTION COMMISSION" if len(l.split()) >= 2 and not re.search(r"election|commission|identity|card|ভোট|নির্বাচন", l, re.I): if _looks_like_name(l): name = l break # Final cleanup: remove stray tokens like "'s Name" if name: name = re.sub(r"(^[\'\"\u2019sS]+\s*|[\'\"\u2019sS]+\s*$)", "", name).strip() if father: father = re.sub(r"(^[\'\"\u2019sS]+\s*|[\'\"\u2019sS]+\s*$)", "", father).strip() fields = { "epic": epic, "name": name if name else None, "father_or_spouse": father if father else None, "dob": dob, "gender": gender, "address": address, } # heuristically compute confidence (simple) filled = sum(1 for v in fields.values() if v) conf = min(0.98, 0.35 + 0.12 * filled) return ExtractResult("voter", fields, conf)