feat(extract): mirror VLM deskew preprocessing into the extract Space

Scanned POs / spec sheets (the high-frequency input routed to hitech-extract) often
arrive tilted, which hurts NuExtract field extraction. Adds _preprocess_image
(Hough-line median-angle deskew, falls back to the raw image on any error) verbatim
from the VLM Space, wires it into _build_messages, and adds opencv-python-headless to
requirements. Closes the long-standing gap (VLM had deskew, extract didn't).

NOTE: takes effect only after a Space redeploy (scripts/sync_spaces.ps1) — HF git is
unreachable from the automated env, so the user runs it. py_compile clean.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.py

@@ -30,7 +30,9 @@ import json
 import re
 from typing import Any
 
+import cv2
 import gradio as gr
+import numpy as np
 import spaces
 import torch
 from PIL import Image
@@ -94,6 +96,51 @@ model = AutoModelForImageTextToText.from_pretrained(
 ).eval()
 
 
+def _preprocess_image(pil_img: Image.Image) -> Image.Image:
+    """Deskew + auto-rotate a scanned document before extraction.
+
+    Scanned POs and spec sheets — the high-frequency input routed to this Space —
+    often arrive slightly tilted from a phone camera or flatbed scanner, which hurts
+    NuExtract's field extraction. Mirrors the VLM Space's preprocessing exactly:
+
+    1. Convert to grayscale and threshold to isolate text/content.
+    2. Find the dominant skew angle via Hough lines and rotate to correct it.
+    3. Return as RGB PIL Image (unchanged if preprocessing fails for any reason).
+    """
+    try:
+        img = np.array(pil_img.convert("RGB"))
+        gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
+        _, binary = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
+
+        # Detect lines to estimate skew angle
+        lines = cv2.HoughLinesP(binary, 1, np.pi / 180, threshold=100, minLineLength=100, maxLineGap=10)
+        if lines is None or len(lines) == 0:
+            return pil_img
+
+        angles = []
+        for line in lines:
+            x1, y1, x2, y2 = line[0]
+            if x2 != x1:
+                angles.append(np.degrees(np.arctan2(y2 - y1, x2 - x1)))
+
+        if not angles:
+            return pil_img
+
+        # Median angle — robust against outlier lines
+        skew = float(np.median(angles))
+        # Only correct small skews (> 0.5° and < 45°) to avoid false rotations
+        if abs(skew) < 0.5 or abs(skew) > 45:
+            return pil_img
+
+        h, w = img.shape[:2]
+        center = (w / 2, h / 2)
+        M = cv2.getRotationMatrix2D(center, skew, 1.0)
+        rotated = cv2.warpAffine(img, M, (w, h), flags=cv2.INTER_LINEAR, borderMode=cv2.BORDER_REPLICATE)
+        return Image.fromarray(rotated)
+    except Exception:
+        return pil_img
+
+
 def _clean_json(text: str) -> str:
     """Drop <think> blocks, code fences and prose; keep the JSON object.
 
@@ -121,7 +168,8 @@ def _clean_json(text: str) -> str:
 def _build_messages(prompt: str, image_path: str | None) -> list[dict]:
     content: list[dict] = []
     if image_path:
-        content.append({"type": "image", "image": Image.open(image_path).convert("RGB")})
+        img = _preprocess_image(Image.open(image_path).convert("RGB"))
+        content.append({"type": "image", "image": img})
     content.append({"type": "text", "text": prompt})
     return [{"role": "user", "content": content}]
 

requirements.txt

@@ -10,3 +10,4 @@ transformers>=4.57.0
 accelerate>=1.0
 torchvision
 pillow
+opencv-python-headless