Spaces:
Running
Running
File size: 3,988 Bytes
06bd65a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 | import io
import cv2
from PIL import Image, ImageOps, ImageFilter
import numpy as np
# Minimum side length below which we upscale the image before sending to OCR.
# Camera photos below this threshold tend to have insufficient pixel density for tiny text.
_MIN_SHORT_SIDE = 1000
class ImageProcessor:
@staticmethod
def process_bytes_to_numpy(image_bytes: bytes) -> np.ndarray:
"""
Converts raw image bytes to an RGB NumPy array ready for PaddleOCR.
Pipeline:
1. EXIF auto-rotation (fixes portrait/landscape mobile photos)
2. Minimum-resolution guard (upscale if image is too small)
3. CLAHE contrast enhancement (equalises shadows / lighting gradients)
4. Unsharp masking (sharpens soft/blurry edges without amplifying noise)
"""
img = Image.open(io.BytesIO(image_bytes))
# 1. Auto-rotate based on EXIF orientation tag (critical for mobile photos)
img = ImageOps.exif_transpose(img)
# Ensure RGB
img_rgb = img.convert("RGB")
# 2. Minimum resolution guard: upscale if the shorter side is too small
img_rgb = ImageProcessor._ensure_min_resolution(img_rgb)
img_np = np.array(img_rgb)
# 3. CLAHE contrast enhancement
try:
img_np = ImageProcessor.enhance_contrast(img_np)
except Exception as e:
print(f"Warning: Contrast enhancement failed ({e}). Proceeding with raw image.")
# 4. Unsharp masking for edge sharpening (applied on PIL image, then back to numpy)
try:
img_np = ImageProcessor._apply_unsharp_mask(img_np)
except Exception as e:
print(f"Warning: Unsharp masking failed ({e}). Skipping sharpening step.")
return img_np
# ------------------------------------------------------------------
# Private helpers
# ------------------------------------------------------------------
@staticmethod
def _ensure_min_resolution(img: Image.Image) -> Image.Image:
"""
If the shorter side of the image is below _MIN_SHORT_SIDE pixels,
scale the image up proportionally using high-quality Lanczos resampling.
This preserves tiny text that would otherwise be unreadable at low DPI.
"""
w, h = img.size
short_side = min(w, h)
if short_side < _MIN_SHORT_SIDE:
scale = _MIN_SHORT_SIDE / short_side
new_w = int(w * scale)
new_h = int(h * scale)
img = img.resize((new_w, new_h), Image.LANCZOS)
return img
@staticmethod
def enhance_contrast(img_np: np.ndarray) -> np.ndarray:
"""
Applies Contrast Limited Adaptive Histogram Equalization (CLAHE)
to the luminance channel to normalise uneven lighting and shadows.
"""
yuv = cv2.cvtColor(img_np, cv2.COLOR_RGB2YUV)
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
yuv[:, :, 0] = clahe.apply(yuv[:, :, 0])
return cv2.cvtColor(yuv, cv2.COLOR_YUV2RGB)
@staticmethod
def _apply_unsharp_mask(img_np: np.ndarray) -> np.ndarray:
"""
Applies a gentle unsharp mask to sharpen character edges.
Parameters tuned conservatively so we enhance detail without
amplifying noise or creating ringing artefacts on fine print.
radius=1.5 — small kernel, only sharpens fine detail
percent=120 — 20% boost in edge contrast
threshold=3 — only sharpen pixels that differ by at least 3 levels
(prevents noise from being sharpened)
"""
# Work in PIL for built-in UnsharpMask
pil_img = Image.fromarray(img_np)
sharpened = pil_img.filter(ImageFilter.UnsharpMask(radius=1.5, percent=120, threshold=3))
return np.array(sharpened)
image_processor = ImageProcessor()
|