Update app.py

app.py

@@ -1,8 +1,8 @@
 """
-GLM-OCR Hugging Face Space — MaaS mode with header extraction.
-Header is extracted from PDF text layer using bbox gap detection.
-Footer is NOT extracted separately — API already captures it as text regions.
+GLM-OCR Hugging Face Space app with client-side header/footer fallback for MaaS.
+Works for every PDF and every image: uses API bboxes when available, else minimal band.
 """
+# Patch asyncio first (before Gradio imports it) to suppress Python 3.13 cleanup noise
 import asyncio
 try:
     _orig_close = asyncio.BaseEventLoop.close
@@ -21,44 +21,79 @@ import re
 import tempfile
 import yaml
 import gradio as gr
+
 import glmocr
 
 log = logging.getLogger("glmocr_app")
 
-GLMOCR_BASE    = os.path.dirname(glmocr.__file__)
-CONFIG_PATH    = os.path.join(GLMOCR_BASE, "config.yaml")
+GLMOCR_BASE = os.path.dirname(glmocr.__file__)
+CONFIG_PATH = os.path.join(GLMOCR_BASE, "config.yaml")
 FORMATTER_PATH = os.path.join(GLMOCR_BASE, "postprocess", "result_formatter.py")
 
-DEFAULT_HEADER_FRAC = float(os.environ.get("GLMOCR_DEFAULT_HEADER_FRAC", "0.12"))
+DEFAULT_ZONE_FRAC = float(os.environ.get("GLMOCR_DEFAULT_ZONE_FRAC", "0.12"))
+MIN_CROP_HEIGHT = int(os.environ.get("GLMOCR_MIN_CROP_HEIGHT", "112"))
+MIN_CROP_PIXELS = int(os.environ.get("GLMOCR_MIN_CROP_PIXELS", "12544"))
+PDF_HEADER_BAND_FRAC = float(os.environ.get("GLMOCR_PDF_HEADER_BAND", "0.15"))
+
+# Single shared parser
+_parser = None
+
+def get_parser():
+    global _parser
+    if _parser is None:
+        from glmocr import GlmOcr
+        _parser = GlmOcr(
+            api_key=os.environ.get("GLMOCR_API_KEY", "4570c28bdea5493c9efae9dae68edc66.sGbA9DLlcX1GlvqV"),
+            mode="maas",
+        )
+    return _parser
 
-# ── STEP 1: Fix config ────────────────────────────────────────
+# ---------------------------------------------------------------------------
+# 1. Config
+# ---------------------------------------------------------------------------
 try:
     with open(CONFIG_PATH, "r") as f:
         config = yaml.safe_load(f)
-
     config["pipeline"]["maas"]["enabled"] = True
-    config["pipeline"]["maas"]["api_key"] = os.environ.get(
-        "GLMOCR_API_KEY", "4570c28bdea5493c9efae9dae68edc66.sGbA9DLlcX1GlvqV"
-    )
+    config["pipeline"]["maas"]["api_key"] = os.environ.get("GLMOCR_API_KEY", "4570c28bdea5493c9efae9dae68edc66.sGbA9DLlcX1GlvqV")
+
+    to_include_as_text = {"header", "footer"}
+    layout_section = config.get("pipeline", {}).get("layout", {})
+    if "label_task_mapping" in layout_section:
+        mapping = layout_section["label_task_mapping"]
+        abandon = mapping.get("abandon")
+        if isinstance(abandon, list):
+            mapping["abandon"] = [x for x in abandon if x not in to_include_as_text]
+        text_labels = mapping.get("text")
+        if isinstance(text_labels, list):
+            for label in to_include_as_text:
+                if label not in text_labels:
+                    text_labels.append(label)
 
-    to_include = {"header", "footer"}
     formatter_section = config.get("pipeline", {}).get("result_formatter", {})
-    if "abandon" in formatter_section and isinstance(formatter_section["abandon"], list):
-        formatter_section["abandon"] = [
-            x for x in formatter_section["abandon"] if x not in to_include
-        ]
+    if "abandon" in formatter_section:
+        abandon = formatter_section["abandon"]
+        if isinstance(abandon, list):
+            formatter_section["abandon"] = [x for x in abandon if x not in to_include_as_text]
+    if "label_visualization_mapping" in formatter_section:
+        text_vis = formatter_section["label_visualization_mapping"].get("text")
+        if isinstance(text_vis, list):
+            for label in to_include_as_text:
+                if label not in text_vis:
+                    text_vis.append(label)
 
     with open(CONFIG_PATH, "w") as f:
         yaml.dump(config, f, default_flow_style=False, sort_keys=False)
 except Exception:
     pass
 
-# ── STEP 2: Fix result_formatter.py ──────────────────────────
+# ---------------------------------------------------------------------------
+# 2. result_formatter
+# ---------------------------------------------------------------------------
 try:
     with open(FORMATTER_PATH, "r") as f:
         source = f.read()
-    for label in ('"header"', "'header'", '"footer"', "'footer'",
-                  '"doc_header"', "'doc_header'", '"doc_footer"', "'doc_footer'"):
+    for label in ('"header"', "'header'", '"footer"', "'footer'", '"doc_header"', "'doc_header'", '"doc_footer"', "'doc_footer'"):
         source = re.sub(r",\s*" + re.escape(label), "", source)
         source = re.sub(re.escape(label) + r"\s*,", "", source)
         source = re.sub(re.escape(label), "", source)
@@ -67,25 +102,14 @@ try:
 except Exception:
     pass
 
-# ── Single shared parser ─────────────────────────────���────────
-_parser = None
 
-def get_parser():
-    global _parser
-    if _parser is None:
-        from glmocr import GlmOcr
-        _parser = GlmOcr(
-            api_key=os.environ.get("GLMOCR_API_KEY", "4570c28bdea5493c9efae9dae68edc66.sGbA9DLlcX1GlvqV"),
-            mode="maas",
-        )
-    return _parser
-
-# ── STEP 3: Header extraction helpers ────────────────────────
 def get_top_gap_frac(regions, img_height):
     """
-    Find the fraction of image height that the API missed at the top.
-    Uses bbox_2d of the first (topmost) region.
-    Returns a fraction (0–1) or None if no gap detected.
+    Find only the TOP gap the API missed using bbox_2d.
+    Returns fraction of image height, or None if no gap.
+    We never extract the bottom zone — the API already captures
+    footer content as text regions, and extracting bottom zone
+    causes duplication of body content on pages without real footers.
     """
     y_tops = []
     for r in regions:
@@ -93,96 +117,166 @@ def get_top_gap_frac(regions, img_height):
         if bbox and len(bbox) >= 4:
             y_tops.append(bbox[1])
     if not y_tops:
-        return DEFAULT_HEADER_FRAC  # no bbox info — use default band
+        return DEFAULT_ZONE_FRAC  # no bbox info — use default band
     first_y = min(y_tops)
-    # Only treat as missed header if gap > 8% of image height
     return (first_y / img_height) if first_y > img_height * 0.08 else None
 
 
-def extract_header_text(pdf_path, page_num, y_end_frac):
-    """Extract text from the top zone of a PDF page using PyMuPDF."""
+def extract_zone_text_pdf(pdf_path, page_num, y_start_frac, y_end_frac):
     try:
         import pymupdf as fitz
-        doc  = fitz.open(pdf_path)
+        doc = fitz.open(pdf_path)
         page = doc[page_num]
         h, w = page.rect.height, page.rect.width
-        text = page.get_text(clip=fitz.Rect(0, 0, w, h * y_end_frac)).strip()
+        rect = fitz.Rect(0, h * y_start_frac, w, h * y_end_frac)
+        text = page.get_text(clip=rect).strip()
         doc.close()
         return text
     except Exception:
         return ""
 
 
-def get_page_data(page_result):
-    """Extract markdown and regions from one PipelineResult."""
+def extract_pdf_text_in_band(pdf_path, page_num, y_start_frac, y_end_frac):
+    try:
+        import pymupdf as fitz
+        doc = fitz.open(pdf_path)
+        page = doc[page_num]
+        h = page.rect.height
+        y_lo = h * y_start_frac
+        y_hi = h * y_end_frac
+        words = page.get_text("words")
+        doc.close()
+        parts = []
+        for w in words:
+            if len(w) >= 5:
+                y0, y1 = float(w[1]), float(w[3])
+                if y0 < y_hi and y1 > y_lo:
+                    parts.append(w[4])
+        return " ".join(parts).strip()
+    except Exception:
+        return ""
+
+
+def ocr_zone(image_path, y_start_frac, y_end_frac):
+    """Run OCR on a horizontal band — used only for header on image files (non-PDF)."""
+    try:
+        from PIL import Image
+        img = Image.open(image_path).convert("RGB")
+        w, h = img.size
+        y0 = max(0, int(h * y_start_frac))
+        y1 = min(h, int(h * y_end_frac))
+        if y1 <= y0:
+            return ""
+        crop = img.crop((0, y0, w, y1))
+        cw, ch = crop.size
+        if ch < MIN_CROP_HEIGHT or (cw * ch) < MIN_CROP_PIXELS:
+            need_h = max(ch, MIN_CROP_HEIGHT)
+            need_w = max(cw, 1)
+            if (need_w * need_h) < MIN_CROP_PIXELS:
+                need_w = max(need_w, (MIN_CROP_PIXELS + need_h - 1) // need_h)
+            canvas = Image.new("RGB", (need_w, need_h), (255, 255, 255))
+            canvas.paste(crop, (0, 0))
+            crop = canvas
+        fd, path = tempfile.mkstemp(suffix=".jpg")
+        os.close(fd)
+        try:
+            crop.save(path, "JPEG", quality=92)
+            parser = get_parser()
+            out = parser.parse(path)
+            if not isinstance(out, list):
+                out = [out]
+            if out and getattr(out[0], "markdown_result", None):
+                return (out[0].markdown_result or "").strip()
+        finally:
+            try:
+                os.unlink(path)
+            except Exception:
+                pass
+    except Exception as e:
+        log.warning("[header] ocr_zone failed: %s", e, exc_info=True)
+    return ""
+
+
+def get_page_md_and_regions(page_result):
     md = ""
     if hasattr(page_result, "markdown_result") and page_result.markdown_result:
         md = (page_result.markdown_result or "").strip()
     regions = []
     if hasattr(page_result, "json_result"):
         jr = page_result.json_result
-        if isinstance(jr, list) and len(jr) > 0:
+        if isinstance(jr, dict) and "regions" in jr:
+            regions = jr.get("regions") or []
+        elif isinstance(jr, list) and len(jr) > 0:
             r = jr[0] if isinstance(jr[0], list) else jr
             if isinstance(r, list):
                 regions = r
+            elif isinstance(r, dict) and "regions" in r:
+                regions = r.get("regions") or []
     return md, regions
 
 
-# ── STEP 4: Main OCR function ─────────────────────────────────
 def run_ocr(uploaded_file):
     if uploaded_file is None:
         return "Please upload a file."
     try:
         import pymupdf as fitz
 
-        path   = uploaded_file.name if hasattr(uploaded_file, "name") else str(uploaded_file)
+        path = uploaded_file.name if hasattr(uploaded_file, "name") else str(uploaded_file)
         is_pdf = path.lower().endswith(".pdf")
         parser = get_parser()
 
         if is_pdf:
-            doc          = fitz.open(path)
-            page_images  = []
+            doc = fitz.open(path)
+            page_images = []
             page_heights = []
             for i in range(len(doc)):
-                pix      = doc[i].get_pixmap(matrix=fitz.Matrix(1.5, 1.5), alpha=False)
+                pix = doc[i].get_pixmap(matrix=fitz.Matrix(1.5, 1.5), alpha=False)
                 img_path = os.path.join(tempfile.gettempdir(), f"maas_page_{i}.png")
                 pix.save(img_path)
                 page_images.append(img_path)
-                # Image height at 1.5x matches bbox_2d coordinate space
                 page_heights.append(doc[i].rect.height * 1.5)
             doc.close()
             results = parser.parse(page_images)
         else:
-            page_images  = [path]
+            page_images = [path]
             page_heights = []
-            results      = parser.parse(path)
+            results = parser.parse(path)
 
         if not isinstance(results, list):
             results = [results]
 
         all_pages = []
-
         for page_num, page_result in enumerate(results):
-            page_md, regions = get_page_data(page_result)
+            page_md, regions = get_page_md_and_regions(page_result)
             parts = []
 
             # ── HEADER at TOP ─────────────────────────────────
-            # Only for PDFs — extract top zone the API missed.
-            # We do NOT extract footer — the API already captures
-            # footer text (copyright, address) as regular text regions.
-            if is_pdf and page_num < len(page_heights):
-                top_gap = get_top_gap_frac(regions, page_heights[page_num])
-                if top_gap is not None:
-                    hdr = extract_header_text(path, page_num, top_gap)
-                    if hdr:
-                        parts.append(hdr)
+            if page_num < len(page_images):
+                img_path = page_images[page_num]
+                hdr = ""
+                if is_pdf and page_num < len(page_heights):
+                    top_gap = get_top_gap_frac(regions, page_heights[page_num])
+                    if top_gap is not None:
+                        hdr = extract_zone_text_pdf(path, page_num, 0, top_gap)
+                        if not hdr:
+                            hdr = extract_pdf_text_in_band(path, page_num, 0, PDF_HEADER_BAND_FRAC)
+                else:
+                    # Image input — use OCR on top band
+                    top_gap = get_top_gap_frac(regions, 1000)
+                    if top_gap is not None:
+                        hdr = ocr_zone(img_path, 0, top_gap)
+                if hdr:
+                    parts.append(hdr)
 
             # ── BODY in MIDDLE ────────────────────────────────
             if page_md:
                 parts.append(page_md)
 
-            # No footer extraction — avoids duplicating body content.
-            # The API captures real footer text as text regions in page_md.
+            # ── NO footer extraction ──────────────────────────
+            # The API already captures real footer text (copyright,
+            # address etc.) as text regions inside page_md.
+            # Extracting the bottom zone causes duplication on pages
+            # where the last body content sits near the page bottom.
 
             if parts:
                 all_pages.append("\n\n".join(parts))
@@ -195,15 +289,11 @@ def run_ocr(uploaded_file):
         return f"Error: {e}\n\n{traceback.format_exc()}"
 
 
-# ── STEP 5: Gradio UI ─────────────────────────────────────────
 with gr.Blocks(title="GLM-OCR") as demo:
-    gr.Markdown("# 🔍 GLM-OCR\nUpload a PDF or image. Headers included in correct position.")
-    file_in = gr.File(
-        label="Upload PDF or image",
-        file_types=[".pdf", ".png", ".jpg", ".jpeg", ".tiff", ".bmp"]
-    )
-    run_btn = gr.Button("▶  Run OCR", variant="primary")
-    out     = gr.Textbox(lines=40, label="Output")
+    gr.Markdown("# GLM-OCR\nUpload a PDF or image. Headers and footers included.")
+    file_in = gr.File(label="Upload PDF or image", file_types=[".pdf", ".png", ".jpg", ".jpeg", ".tiff", ".bmp"])
+    run_btn = gr.Button("Run OCR", variant="primary")
+    out = gr.Textbox(lines=40, label="Output")
     run_btn.click(fn=run_ocr, inputs=file_in, outputs=out)
 
 if __name__ == "__main__":