Update app.py

app.py

@@ -1,6 +1,8 @@
 """
-GLM-OCR Hugging Face Space app with client-side header fallback for MaaS.
+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
@@ -30,8 +32,9 @@ FORMATTER_PATH = os.path.join(GLMOCR_BASE, "postprocess", "result_formatter.py")
 
 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"))
+MIN_CROP_PIXELS = int(os.environ.get("GLMOCR_MIN_CROP_PIXELS", "12544"))  # 112*112
+PDF_HEADER_BAND_FRAC = float(os.environ.get("GLMOCR_PDF_HEADER_BAND", "0.15"))   # top 15%
+PDF_FOOTER_BAND_FRAC = float(os.environ.get("GLMOCR_PDF_FOOTER_BAND", "0.85"))   # bottom 15%
 
 _parser = None
 
@@ -46,7 +49,7 @@ def get_parser():
     return _parser
 
 # ---------------------------------------------------------------------------
-# 1. Config
+# 1. Config: set MaaS and optionally include headers/footers for non-MaaS
 # ---------------------------------------------------------------------------
 try:
     with open(CONFIG_PATH, "r") as f:
@@ -82,10 +85,10 @@ try:
     with open(CONFIG_PATH, "w") as f:
         yaml.dump(config, f, default_flow_style=False, sort_keys=False)
 except Exception:
-    pass
+    pass  # e.g. read-only package on Hugging Face; MaaS ignores layout anyway
 
 # ---------------------------------------------------------------------------
-# 2. result_formatter
+# 2. result_formatter: stop stripping header/footer (best-effort; may be read-only)
 # ---------------------------------------------------------------------------
 try:
     with open(FORMATTER_PATH, "r") as f:
@@ -100,21 +103,22 @@ except Exception:
     pass
 
 
-def get_header_zone(regions, norm_height=1000):
-    """Return only the top gap fraction. Never touch the bottom."""
+def get_header_footer_zones(regions, norm_height=1000):
     if not regions:
-        return None
-    y_tops = []
+        return None, None
+    y_tops, y_bottoms = [], []
     for r in regions:
         bbox = r.get("bbox_2d") if isinstance(r, dict) else getattr(r, "bbox_2d", None)
         if bbox and len(bbox) >= 4:
             y_tops.append(bbox[1])
+            y_bottoms.append(bbox[3])
     if not y_tops:
-        return None
-    return min(y_tops) / norm_height
+        return None, None
+    return min(y_tops) / norm_height, max(y_bottoms) / norm_height
 
 
 def extract_zone_text_pdf(pdf_path, page_num, y_start_frac, y_end_frac):
+    """Extract text from a horizontal band using a clip rect."""
     try:
         import pymupdf as fitz
         doc = fitz.open(pdf_path)
@@ -129,6 +133,7 @@ def extract_zone_text_pdf(pdf_path, page_num, y_start_frac, y_end_frac):
 
 
 def extract_pdf_text_in_band(pdf_path, page_num, y_start_frac, y_end_frac):
+    """Extract all text whose word bbox falls in the given vertical band (by position)."""
     try:
         import pymupdf as fitz
         doc = fitz.open(pdf_path)
@@ -136,7 +141,7 @@ def extract_pdf_text_in_band(pdf_path, page_num, y_start_frac, y_end_frac):
         h = page.rect.height
         y_lo = h * y_start_frac
         y_hi = h * y_end_frac
-        words = page.get_text("words")
+        words = page.get_text("words")  # list of (x0, y0, x1, y1, word, ...)
         doc.close()
         parts = []
         for w in words:
@@ -150,6 +155,8 @@ def extract_pdf_text_in_band(pdf_path, page_num, y_start_frac, y_end_frac):
 
 
 def ocr_zone(image_path, y_start_frac, y_end_frac):
+    """Run OCR on a horizontal band. Pads small crops to meet API minimum size."""
+    zone_name = "header" if y_end_frac < 0.5 else "footer"
     try:
         from PIL import Image
         img = Image.open(image_path).convert("RGB")
@@ -166,8 +173,12 @@ def ocr_zone(image_path, y_start_frac, y_end_frac):
             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))
+            if zone_name == "header":
+                canvas.paste(crop, (0, 0))
+            else:
+                canvas.paste(crop, (0, need_h - ch))
             crop = canvas
+            cw, ch = crop.size
         fd, path = tempfile.mkstemp(suffix=".jpg")
         os.close(fd)
         try:
@@ -177,14 +188,15 @@ def ocr_zone(image_path, y_start_frac, y_end_frac):
             if not isinstance(out, list):
                 out = [out]
             if out and getattr(out[0], "markdown_result", None):
-                return (out[0].markdown_result or "").strip()
+                text = (out[0].markdown_result or "").strip()
+                return text
         finally:
             try:
                 os.unlink(path)
             except Exception:
                 pass
     except Exception as e:
-        log.warning("[header] ocr_zone failed: %s", e, exc_info=True)
+        log.warning("[%s] ocr_zone failed: %s", zone_name, e, exc_info=True)
     return ""
 
 
@@ -236,14 +248,18 @@ def run_ocr(uploaded_file):
         all_pages = []
         for page_num, page_result in enumerate(results):
             page_md, regions = get_page_md_and_regions(page_result)
+            header_end_frac, footer_start_frac = get_header_footer_zones(regions, 1000)
 
-            he = get_header_zone(regions, 1000)
-            if he is None or he <= 0:
+            he = header_end_frac if header_end_frac is not None else DEFAULT_ZONE_FRAC
+            fs = footer_start_frac if footer_start_frac is not None else (1.0 - DEFAULT_ZONE_FRAC)
+            if he <= 0:
                 he = DEFAULT_ZONE_FRAC
+            if fs >= 1.0:
+                fs = 1.0 - DEFAULT_ZONE_FRAC
 
             parts = []
 
-            # ── HEADER at TOP only ────────────────────────────
+            # Always run header band (top 8-12% of page)
             if page_num < len(page_images):
                 img_path = page_images[page_num]
                 hdr = ""
@@ -256,13 +272,21 @@ def run_ocr(uploaded_file):
                 if hdr and hdr.strip():
                     parts.append(hdr.strip())
 
-            # ── BODY in MIDDLE ────────────────────────────────
             if page_md:
                 parts.append(page_md)
 
-            # ── NO footer extraction ──────────────────────────
-            # The API already captures real footer text correctly.
-            # Extracting bottom zone duplicates body content.
+            # Always run footer band (bottom 8-12% of page)
+            if page_num < len(page_images):
+                img_path = page_images[page_num]
+                ftr = ""
+                if is_pdf:
+                    ftr = extract_zone_text_pdf(path, page_num, fs, 1.0)
+                    if not (ftr and ftr.strip()):
+                        ftr = extract_pdf_text_in_band(path, page_num, PDF_FOOTER_BAND_FRAC, 1.0)
+                if not (ftr and ftr.strip()):
+                    ftr = ocr_zone(img_path, fs, 1.0)
+                if ftr and ftr.strip():
+                    parts.append(ftr.strip())
 
             if parts:
                 all_pages.append("\n\n".join(parts))