Text Extraction Done, UI Fixes Done, Extracted Template Layout Saved in DB

This commit is contained in:
2026-07-12 14:24:17 +05:30
parent 3163bb213e
commit 3d6614f408
79 changed files with 3139 additions and 81 deletions

View File

@@ -0,0 +1,217 @@
import io
import re
from typing import List, Dict, Any
import numpy as np
import cv2
import pdfplumber
import pytesseract
from PIL import Image
from pdf2image import convert_from_bytes
from sqlalchemy.orm import Session
from models.template_models import DocumentLayout, TemplateDocument
from schemas.template_schemas import DocumentLayoutSchema
class DocumentProcessor:
def __init__(self, db: Session):
self.db = db
def process_file(self, document_id: int, filename: str, file_bytes: bytes) -> List[DocumentLayout]:
document = self.db.query(TemplateDocument).filter(TemplateDocument.pk_document_id == document_id).first()
if not document:
raise ValueError("Document not found")
layouts = []
if filename.lower().endswith('.pdf'):
layouts = self._process_pdf(document_id, file_bytes)
elif filename.lower().endswith(('.png', '.jpg', '.jpeg', '.tiff')):
layouts = self._process_image(document_id, file_bytes)
# Update page count
if layouts:
document.page_count = max([l.page_no for l in layouts])
document.status = "processed"
self.db.commit()
return layouts
def _process_pdf(self, document_id: int, file_bytes: bytes) -> List[DocumentLayout]:
layouts = []
try:
with pdfplumber.open(io.BytesIO(file_bytes)) as pdf:
sequence_no = 1
for page_no, page in enumerate(pdf.pages, start=1):
# Check if it has text
text = page.extract_text()
if text and text.strip():
# Extract words with bounding boxes
words = page.extract_words()
# Group words into blocks using simple proximity heuristics
blocks = self._group_words_to_blocks(words, page.width, page.height)
for block in blocks:
db_layout = DocumentLayout(
fk_document_id=document_id,
page_no=page_no,
text_value=block['text'],
block_type=block['type'],
x_coordinate=block['x0'],
y_coordinate=block['top'],
width=block['x1'] - block['x0'],
height=block['bottom'] - block['top'],
confidence=100.0,
sequence_no=sequence_no
)
self.db.add(db_layout)
layouts.append(db_layout)
sequence_no += 1
else:
# Scanned PDF page -> convert to image and process
# Handled separately to avoid complexity in this mock
pass
except Exception as e:
print(f"Error processing PDF with pdfplumber: {e}")
if not layouts:
# Fallback to image-based processing for scanned PDFs
images = convert_from_bytes(file_bytes, dpi=300)
sequence_no = 1
for page_no, img in enumerate(images, start=1):
img_byte_arr = io.BytesIO()
img.save(img_byte_arr, format='PNG')
page_layouts = self._process_image(document_id, img_byte_arr.getvalue(), page_no=page_no, start_sequence=sequence_no)
layouts.extend(page_layouts)
sequence_no += len(page_layouts)
self.db.commit()
return layouts
def _process_image(self, document_id: int, file_bytes: bytes, page_no: int = 1, start_sequence: int = 1) -> List[DocumentLayout]:
layouts = []
try:
# Decode image
nparr = np.frombuffer(file_bytes, np.uint8)
img = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
# Preprocessing
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# Denoise & Threshold
gray = cv2.medianBlur(gray, 3)
gray = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1]
# OCR using PyTesseract with bounding boxes
data = pytesseract.image_to_data(gray, output_type=pytesseract.Output.DICT)
sequence_no = start_sequence
# Grouping words into lines/blocks simplified
height, width = gray.shape
blocks = self._tesseract_data_to_blocks(data, width, height)
for block in blocks:
db_layout = DocumentLayout(
fk_document_id=document_id,
page_no=page_no,
text_value=block['text'],
block_type=block['type'],
x_coordinate=block['x'],
y_coordinate=block['y'],
width=block['w'],
height=block['h'],
confidence=block['conf'],
sequence_no=sequence_no
)
self.db.add(db_layout)
layouts.append(db_layout)
sequence_no += 1
except Exception as e:
print(f"Error processing image with tesseract: {e}")
self.db.commit()
return layouts
def _group_words_to_blocks(self, words: List[Dict], page_width: float, page_height: float) -> List[Dict]:
"""Group nearby words into logical blocks and classify them."""
# Extremely simplified grouping by Y coordinate proximity
blocks = []
if not words:
return blocks
# Sort words by Y then X
words.sort(key=lambda w: (w['top'], w['x0']))
current_block = {
'text': words[0]['text'],
'x0': words[0]['x0'],
'top': words[0]['top'],
'x1': words[0]['x1'],
'bottom': words[0]['bottom']
}
for word in words[1:]:
# If word is roughly on the same line (y diff is small) and close horizontally
if abs(word['top'] - current_block['top']) < 10 and (word['x0'] - current_block['x1']) < 50:
current_block['text'] += ' ' + word['text']
current_block['x1'] = word['x1']
current_block['top'] = min(current_block['top'], word['top'])
current_block['bottom'] = max(current_block['bottom'], word['bottom'])
else:
current_block['type'] = self._classify_block(current_block, page_height)
blocks.append(current_block)
current_block = {
'text': word['text'],
'x0': word['x0'],
'top': word['top'],
'x1': word['x1'],
'bottom': word['bottom']
}
current_block['type'] = self._classify_block(current_block, page_height)
blocks.append(current_block)
return blocks
def _tesseract_data_to_blocks(self, data: Dict, page_width: float, page_height: float) -> List[Dict]:
blocks = []
n_boxes = len(data['text'])
current_line = []
for i in range(n_boxes):
if int(data['conf'][i]) > 10 and data['text'][i].strip():
# We can group by line_num
blocks.append({
'text': data['text'][i],
'x': data['left'][i],
'y': data['top'][i],
'w': data['width'][i],
'h': data['height'][i],
'conf': float(data['conf'][i]),
'type': self._classify_block({'top': data['top'][i], 'text': data['text'][i]}, page_height)
})
return blocks
def _classify_block(self, block: Dict, page_height: float) -> str:
y = block.get('top') or block.get('y') or 0
text = block.get('text', '').lower()
if y < page_height * 0.15:
return "HEADER"
elif y > page_height * 0.85:
return "FOOTER"
if "total" in text:
return "TOTAL"
if "tax" in text or "gst" in text or "vat" in text:
return "TAX"
if re.search(r'\b(vendor|from)\b', text):
return "VENDOR"
if re.search(r'\b(bill to|sold to)\b', text):
return "BILL_TO"
if re.search(r'\b(ship to)\b', text):
return "SHIP_TO"
# Very naive fallback for table check
if re.search(r'\b(qty|rate|amount|price|item)\b', text):
return "TABLE_HEADER"
return "TEXT"

View File

@@ -0,0 +1,174 @@
import hashlib
import json
from typing import List, Dict, Any, Tuple
from sqlalchemy.orm import Session
from models.template_models import Template, TemplateDocument, DocumentLayout, TemplateRecognitionHistory
from schemas.template_schemas import TemplateRecognitionResult
class TemplateRecognitionEngine:
def __init__(self, db: Session):
self.db = db
def generate_fingerprint(self, layouts: List[DocumentLayout]) -> str:
"""
Generate a structural fingerprint based on the document's layout.
We use relative positions of HEADER, VENDOR, TABLE_HEADER blocks.
"""
key_blocks = []
for layout in layouts:
if layout.block_type in ["HEADER", "VENDOR", "TABLE_HEADER", "TOTAL"]:
key_blocks.append({
"type": layout.block_type,
"text": layout.text_value[:50] if layout.text_value else "", # first 50 chars
"rx": round(float(layout.x_coordinate) / 100) if layout.x_coordinate else 0, # relative bucket x
"ry": round(float(layout.y_coordinate) / 100) if layout.y_coordinate else 0 # relative bucket y
})
# Sort top to bottom, left to right
key_blocks.sort(key=lambda b: (b['ry'], b['rx']))
fingerprint_data = json.dumps(key_blocks)
return hashlib.sha256(fingerprint_data.encode('utf-8')).hexdigest()
def recognize_template(self, document_id: int) -> TemplateRecognitionResult:
"""
Matches a document against existing templates using a 4-level weighted score:
Level 1: Vendor Match (40%)
Level 2: Header Similarity (20%)
Level 3: Layout/Fingerprint Similarity (20%)
Level 4: Coordinate Similarity (20%)
Threshold: 85%
"""
layouts = self.db.query(DocumentLayout).filter(DocumentLayout.fk_document_id == document_id).all()
if not layouts:
return TemplateRecognitionResult(templateMatched=False)
doc_fingerprint = self.generate_fingerprint(layouts)
# Extract features for scoring
doc_vendor = self._extract_vendor_name(layouts)
doc_headers = self._extract_headers(layouts)
templates = self.db.query(Template).filter(Template.active_flag == True).all()
best_match = None
highest_score = 0
for template in templates:
score = 0.0
# Level 1: Fingerprint Exact Match (counts for Layout + Vendor + Header if exact)
if template.template_fingerprint == doc_fingerprint:
score += 100.0
else:
# Need to load template sample mapping to compare heuristics
# In a real system, we'd compare against the mapped fields' coordinates
score += self._calculate_heuristic_score(template, doc_vendor, doc_headers, layouts)
if score > highest_score:
highest_score = score
best_match = template
# Store history
if best_match:
history = TemplateRecognitionHistory(
fk_template_id=best_match.pk_template_id,
fk_document_id=document_id,
recognition_score=highest_score,
matched_flag=(highest_score >= 85.0)
)
self.db.add(history)
self.db.commit()
if best_match and highest_score >= 85.0:
extracted_fields = self._auto_map_fields(best_match, layouts)
return TemplateRecognitionResult(
templateMatched=True,
templateId=best_match.pk_template_id,
confidence=highest_score,
extractedFields=extracted_fields
)
return TemplateRecognitionResult(templateMatched=False)
def _extract_vendor_name(self, layouts: List[DocumentLayout]) -> str:
for layout in layouts:
if layout.block_type == "VENDOR":
return layout.text_value.lower()
return ""
def _extract_headers(self, layouts: List[DocumentLayout]) -> List[str]:
headers = []
for layout in layouts:
if layout.block_type == "TABLE_HEADER" and layout.text_value:
headers.append(layout.text_value.lower())
return headers
def _calculate_heuristic_score(self, template: Template, doc_vendor: str, doc_headers: List[str], layouts: List[DocumentLayout]) -> float:
score = 0.0
# This requires the template to have some stored metadata or we check its fields
# E.g., if template has a field "Vendor Name" mapped to a specific text
# 1. Vendor Match (40%)
# For this mockup, we check if the template name matches the vendor
if template.template_name.lower() in doc_vendor or doc_vendor in template.template_name.lower():
score += 40.0
# 2. Header Similarity (20%)
# Check if template fields exist that match doc_headers
field_labels = [f.field_label.lower() for f in template.fields]
header_matches = sum(1 for h in doc_headers if any(h in fl or fl in h for fl in field_labels))
if len(doc_headers) > 0:
score += (header_matches / len(doc_headers)) * 20.0
# 3. Layout / Coordinate Similarity
# We can check the template's previous mappings coordinates against current document
# If coordinates are within a tolerance, we add score
mapping_count = 0
match_count = 0
for mapping in template.mappings:
mapping_count += 1
# Find a layout block in current document that is near mapping coordinates
for layout in layouts:
if layout.page_no == mapping.page_no:
# check distance
if mapping.x_coordinate and mapping.y_coordinate and layout.x_coordinate and layout.y_coordinate:
dx = abs(float(mapping.x_coordinate) - float(layout.x_coordinate))
dy = abs(float(mapping.y_coordinate) - float(layout.y_coordinate))
if dx < 50 and dy < 20: # arbitrary tolerance
match_count += 1
break
if mapping_count > 0:
coord_score = (match_count / mapping_count) * 40.0 # Layout 20% + Coord 20%
score += coord_score
return score
def _auto_map_fields(self, template: Template, layouts: List[DocumentLayout]) -> List[dict]:
extracted = []
for mapping in template.mappings:
# Find closest block in new document
best_block = None
min_dist = float('inf')
for layout in layouts:
if layout.page_no == mapping.page_no:
if mapping.x_coordinate and mapping.y_coordinate and layout.x_coordinate and layout.y_coordinate:
dx = abs(float(mapping.x_coordinate) - float(layout.x_coordinate))
dy = abs(float(mapping.y_coordinate) - float(layout.y_coordinate))
dist = dx**2 + dy**2
if dist < min_dist and dist < 5000: # tolerance squared
min_dist = dist
best_block = layout
if best_block:
extracted.append({
"field_id": mapping.fk_template_field_id,
"field_label": mapping.field.field_label if mapping.field else "",
"value": best_block.text_value,
"confidence": 95.0, # arbitrary high confidence for coordinate match
"layout_id": best_block.pk_document_data_id
})
return extracted