# ------------------- Tables ------------------- # def extract_tables(pdf_path: Path, out_dir: Path): """ Uses tabula-py (Java) to pull out tables. Each table is saved as CSV under out_dir/tables/page_XX_table_YY.csv . """ try: import tabula except ImportError: print("⚠️ tabula-py not installed – skipping table extraction.") return
# ------------------- Main driver ------------------- # def main(): parser = argparse.ArgumentParser( description="Extract a suite of features from a PDF (e.g. agnibina.pdf)." ) parser.add_argument("pdf", type=Path, help="Path to the input PDF") parser.add_argument( "-o", "--out
# ------------------- OCR (optional) ------------------- # def run_ocr_if_needed(pdf_path: Path, out_dir: Path, force: bool = False): """ If the PDF appears to have no extractable text (e.g. scanned), run OCR. Uses ocrmypdf which adds a text layer while preserving the original appearance. """ try: import ocrmypdf except ImportError: print("⚠️ ocrmypdf not installed – OCR step skipped.") return
safe_mkdir(out_dir / "tables") # tabula can auto-detect tables across the whole doc: tables = tabula.read_pdf(str(pdf_path), pages="all", multiple_tables=True, pandas_options='dtype': str) print(f"📊 Detected len(tables) tables.") for i, df in enumerate(tables, start=1): # Try to infer the page number from the DataFrame's metadata if present # (tabula doesn’t expose page number directly; you can run per-page if you need it) csv_path = out_dir / f"tables/table_i:03d.csv" df.to_csv(csv_path, index=False) print(f" → Saved table i → csv_path") agnibina filetype.pdf
ocr_output = out_dir / "ocr_layered.pdf" print("🖼️ Running OCR (this may take a while)…") ocrmypdf.ocr(str(pdf_path), str(ocr_output), force_ocr=True, deskew=True, language="eng") print(f"🆗 OCR complete → ocr_output")
You can pick and choose which of those you need; the code examples below let you toggle them on/off. | Feature | Recommended Library / CLI | Pros | Cons / Gotchas | |---------|---------------------------|------|----------------| | Basic metadata & text | PyPDF2 , pdfminer.six | Pure‑Python, no external dependencies | Struggles with complex layouts, no OCR | | Robust text + layout | pdfplumber (wraps pdfminer ) | Gives you bounding‑box coordinates, easy table extraction | Slower on huge PDFs | | Tables | tabula-py (Java), camelot | Detects table borders, outputs to DataFrames/CSV | Needs Java (tabula) or Ghostscript (camelot) | | Images & embedded files | pdfminer.six (low‑level), pymupdf (aka fitz ) | Fast, easy extraction of images & attachments | pymupdf is C‑based, needs binary wheels | | Full‑featured OCR | pdf2image + pytesseract , or ocrmypdf | Handles scanned PDFs end‑to‑end | Requires Tesseract OCR + poppler; slower | | Metadata & advanced content | Apache Tika (via tika-python ) | Handles many MIME types, auto‑detects language, OCR via Tesseract | Requires a Java runtime; heavier | | Command‑line quick‑look | exiftool , pdfinfo (poppler), mutool (MuPDF) | Great for batch scripts, no Python needed | Limited to what each tool exposes | | Deep NLP (NER, summarisation) | Hugging Face Transformers ( layoutlmv3 , pdfbert ) | Understands layout‑aware entities | Needs GPU for speed, heavier setup | 3. One‑stop Python script (extract most common features) Below is a single, modular script you can drop into a file called extract_agnibina_features.py . It uses only pure‑Python libraries ( pdfplumber , pymupdf ) plus optional OCR ( ocrmypdf ). Feel free to comment out the sections you don’t need.
img_counter = 0 for page_num in tqdm(range(len(doc)), desc="Pages (images)"): page = doc[page_num] img_list = page.get_images(full=True) for img_index, img in enumerate(img_list, start=1): xref = img[0] base_image = doc.extract_image(xref) img_bytes = base_image["image"] img_ext = base_image["ext"] img_name = f"pagepage_num+1:03d_imgimg_index:03d.img_ext" (img_dir / img_name).write_bytes(img_bytes) img_counter += 1 doc.close() print(f"✅ Extracted img_counter images to img_dir") agnibina
# ------------------- Images ------------------- # def extract_images(pdf_path: Path, out_dir: Path): """Extract every image to out_dir/images/ (preserves original format).""" doc = fitz.open(str(pdf_path)) img_dir = out_dir / "images" safe_mkdir(img_dir)
#!/usr/bin/env python3 # -*- coding: utf-8 -*-
# ------------------- Text + Layout ------------------- # def extract_text_and_layout(pdf_path: Path, out_dir: Path) -> List[Dict]: """ Returns a list (one dict per page) with: - page_number - plain_text - list of text elements text, x0, y0, x1, y1, fontname, size """ pages_info = [] with pdfplumber.open(str(pdf_path)) as pdf: for page_num, page in enumerate(tqdm(pdf.pages, desc="Pages (text/layout)")): plain = page.extract_text() # layout objects (characters) – useful for heading detection chars = page.chars # each char already has x0, y0, x1, y1, fontname, size # Group chars into words/lines if you like, but we keep raw for flexibility pages_info.append( "page_number": page_num + 1, "text": plain, "characters": chars, ) # Save raw JSON for later inspection (out_dir / "text_layout.json").write_text(json.dumps(pages_info, indent=2, ensure_ascii=False)) return pages_info img in enumerate(img_list
# ------------------- Metadata ------------------- # def extract_metadata(pdf_path: Path) -> Dict: """Return a dict with PDF metadata (title, author, dates, etc.).""" doc = fitz.open(str(pdf_path)) meta = doc.metadata # Normalize keys normalized = "title": meta.get("title"), "author": meta.get("author"), "creator": meta.get("creator"), "producer": meta.get("producer"), "subject": meta.get("subject"), "keywords": meta.get("keywords"), "creationDate": meta.get("creationDate"), "modDate": meta.get("modDate"), "pdf_version": doc.pdf_version, "page_count": doc.page_count, doc.close() return normalized
count = 0 for i in range(doc.embfile_count()): info = doc.embfile_info(i) fname = clean_filename(info["filename"]) data = doc.embfile_get(i) (att_dir / fname).write_bytes(data) count += 1 doc.close() print(f"📦 Extracted count embedded file(s).")
Features covered: * Basic metadata * Full text (with page numbers) * Text layout (coordinates, fonts) * Images (saved to disk) * Tables (as CSV) * Bookmarks / outline * Embedded files (attachments) * Optional OCR for scanned PDFs
import pdfplumber import fitz # pymupdf from tqdm import tqdm