page-dewarp 0.3.1

Page dewarping and thresholding using a cubic sheet model.

page-dewarp

Document image dewarping library using a cubic sheet model

Python 3 library for page dewarping and thresholding, available on PyPI.

Installation

To install from PyPI, optionally using uv (recommended), run:

uv pip install page-dewarp

JAX

To install with JAX autodiff for ~11x faster optimisation on single images and ~33x faster for batches (CPU only), add the jax extra:

uv pip install page-dewarp[jax]

GPU support

To install with support for GPU execution instead of only CPU, choose one of:

uv pip install page-dewarp[jax-cuda12] # CUDA 12
uv pip install page-dewarp[jax-cuda13] # CUDA 13 (requires Python 3.11+)

Note: CPU execution is the default DEVICE and can be faster than GPU for this workload, but this may vary depending on your relative CPU/GPU horsepower (cores, RAM, VRAM, etc.)

Serial vs Batch

When the JAX backend is available, default behaviour when given multiple images is to use batch mode. Performance benchmark on 40 images (via #139):

Device	Serial	Batch	Speedup
CPU	36s	8.7s	4.1x
GPU	53s	11.2s	4.7x

Dependencies

Python 3.10+ and NumPy, SciPy, SymPy, Matplotlib, OpenCV, and msgspec are required to run page-dewarp.

Help

See documentation for more details.

• Update!: the docs now have a How It Works section

Background

This library was renovated from the original (2016) Python 2 script by Matt Zucker, as Python 2 is now long since decommissioned.

Usage

• See config docs for a table of options

usage: page-dewarp [-h] [-d {0,1,2,3}] [-dd {file,screen,both}]
                   [-o OUTPUT_DIR] [-it OPT_MAX_ITER] [-m OPT_METHOD]
                   [-dev DEVICE] [-b USE_BATCH] [-vw SCREEN_MAX_W]
                   [-vh SCREEN_MAX_H] [-x PAGE_MARGIN_X] [-y PAGE_MARGIN_Y]
                   [-tw TEXT_MIN_WIDTH] [-th TEXT_MIN_HEIGHT]
                   [-ta TEXT_MIN_ASPECT] [-tk TEXT_MAX_THICKNESS]
                   [-tm TEXT_MORPH_OPS] [-lm LINE_MORPH_OPS]
                   [-wz ADAPTIVE_WINSZ] [-ri RVEC_IDX] [-ti TVEC_IDX]
                   [-ci CUBIC_IDX] [-sw SPAN_MIN_WIDTH] [-sp SPAN_PX_PER_STEP]
                   [-eo EDGE_MAX_OVERLAP] [-el EDGE_MAX_LENGTH]
                   [-ec EDGE_ANGLE_COST] [-ea EDGE_MAX_ANGLE]
                   [-f FOCAL_LENGTH] [-z OUTPUT_ZOOM] [-dpi OUTPUT_DPI]
                   [-nb NO_BINARY] [-sh SHEAR_COST] [-mc MAX_CORR]
                   [-s REMAP_DECIMATE]
                   IMAGE_FILE_OR_FILES [IMAGE_FILE_OR_FILES ...]

positional arguments:
  IMAGE_FILE_OR_FILES   One or more images to process

options:
  -h, --help            show this help message and exit
  -d, --debug-level {0,1,2,3}
                        (type: int, default: 0)
  -dd, --debug-dest {file,screen,both}
                        (type: str, default: file)
  -o, --output-dir OUTPUT_DIR
                        Directory for output and debug images (type: str,
                        default: .)
  -it, --max-iter OPT_MAX_ITER
                        Maximum optimisation iterations (type: int, default:
                        600000)
  -m, --method OPT_METHOD
                        Name of the JAX/SciPy optimisation method to use.
                        (type: str, default: auto)
  -dev, --device DEVICE
                        Compute device to select for optimisation. (type: str,
                        default: auto)
  -b, --batch USE_BATCH
                        Whether to batch process images (JAX backend only).
                        (type: str, default: auto)
  -vw, --max-screen-width SCREEN_MAX_W
                        Viewing screen max width (for resizing to screen)
                        (type: int, default: 1280)
  -vh, --max-screen-height SCREEN_MAX_H
                        Viewing screen max height (for resizing to screen)
                        (type: int, default: 700)
  -x, --x-margin PAGE_MARGIN_X
                        Reduced px to ignore near L/R edge (type: int,
                        default: 50)
  -y, --y-margin PAGE_MARGIN_Y
                        Reduced px to ignore near T/B edge (type: int,
                        default: 20)
  -tw, --min-text-width TEXT_MIN_WIDTH
                        Min reduced px width of detected text contour (type:
                        int, default: 15)
  -th, --min-text-height TEXT_MIN_HEIGHT
                        Min reduced px height of detected text contour (type:
                        int, default: 2)
  -ta, --min-text-aspect TEXT_MIN_ASPECT
                        Filter out text contours below this w/h ratio (type:
                        float, default: 1.5)
  -tk, --max-text-thickness TEXT_MAX_THICKNESS
                        Max reduced px thickness of detected text contour
                        (type: int, default: 10)
  -tm, --text-morph TEXT_MORPH_OPS
                        Morphological ops for text mask (e.g. d_9_1,e_1_3)
                        (type: str, default: d_9_1,e_1_3)
  -lm, --line-morph LINE_MORPH_OPS
                        Morphological ops for line mask (e.g. e_3_1_3,d_8_2)
                        (type: str, default: e_3_1_3,d_8_2)
  -wz, --adaptive-winsz ADAPTIVE_WINSZ
                        Window size for adaptive threshold in reduced px
                        (type: int, default: 55)
  -ri, --rotation-vec-param-idx RVEC_IDX
                        Index of rvec in params vector (slice: pair of values)
                        (type: tuple[int, int], default: (0, 3))
  -ti, --translation-vec-param-idx TVEC_IDX
                        Index of tvec in params vector (slice: pair of values)
                        (type: tuple[int, int], default: (3, 6))
  -ci, --cubic-slope-param-idx CUBIC_IDX
                        Index of cubic slopes in params vector (slice: pair of
                        values) (type: tuple[int, int], default: (6, 8))
  -sw, --min-span-width SPAN_MIN_WIDTH
                        Minimum reduced px width for span (type: int, default:
                        30)
  -sp, --span-spacing SPAN_PX_PER_STEP
                        Reduced px spacing for sampling along spans (type:
                        int, default: 20)
  -eo, --max-edge-overlap EDGE_MAX_OVERLAP
                        Max reduced px horiz. overlap of contours in span
                        (type: float, default: 1.0)
  -el, --max-edge-length EDGE_MAX_LENGTH
                        Max reduced px length of edge connecting contours
                        (type: float, default: 100.0)
  -ec, --edge-angle-cost EDGE_ANGLE_COST
                        Cost of angles in edges (tradeoff vs. length) (type:
                        float, default: 10.0)
  -ea, --max-edge-angle EDGE_MAX_ANGLE
                        Maximum change in angle allowed between contours
                        (type: float, default: 7.5)
  -f, --focal-length FOCAL_LENGTH
                        Normalized focal length of camera (type: float,
                        default: 1.2)
  -z, --output-zoom OUTPUT_ZOOM
                        How much to zoom output relative to *original* image
                        (type: float, default: 1.0)
  -dpi, --output-dpi OUTPUT_DPI
                        Just affects stated DPI of PNG, not appearance (type:
                        int, default: 300)
  -nb, --no-binary NO_BINARY
                        Disable output conversion to binary thresholded image
                        (type: int, default: 0)
  -sh, --shear-cost SHEAR_COST
                        Penalty against camera tilt (shear distortion). (type:
                        float, default: 0.0)
  -mc, --max-corrections MAX_CORR
                        Maximum corrections used to approximate the inverse
                        Hessian. (type: int, default: 100)
  -s, --shrink REMAP_DECIMATE
                        Downscaling factor for remapping image (type: int,
                        default: 16)

To try out an example image, run

git clone https://github.com/lmmx/page-dewarp
cd page-dewarp
mkdir results && cd results
page-dewarp ../example_input/boston_cooking_a.jpg

Explanation and further reading

A book on a flat surface can be said to be 'fixed to zero' at the endpoints of a curve, which you can model as a cubic (see derive_cubic.py)

The "cubic Hermite spline" is one of the models supported by Gpufit, a library for Levenberg Marquardt curve fitting in CUDA (C++ with Python API).

[Work in progress]

• See full writeup on Matt Zucker's blog
• See lecture on splines by Steve Marschner for more details and how a spline can be represented in matrix form.
• Brief notes on this project are over on my website

Features

Improvements on the original include:

• Banished Python 2
• Command line interface
  • Alterable config options
• Repackage for pip installation
• Refactor with modules and classes
• Speed up the optimisation
  • Limit optimisation iterations (via -it flag)
  • Optional GPU interface (via -dev flag)
  • Multiprocessing on CPU