foureg 1.0.0

Fourier transform based image registration

Foureg

Image registration using discrete Fourier transform.

Given two images, foureg calculates a similarity transformation that transforms one image into the other.

Example

The example transforms an image with a user defined transformation and then rediscovers it using foureg.

import matplotlib.pyplot as plt
import numpy as np
import torch
from PIL import Image

from foureg import (Constraints, frame_img, similarity, similarity_matrix,
                    transform_img)

# Generate the test images
transformation = similarity_matrix(0.8, -10, (60, 20))
master = np.asarray(Image.open("./resources/examples/sample1.png"))
master = torch.from_numpy(master.copy()).type(torch.float32)

slave = transform_img(master, transformation, invert=True)

# Define some constraints and coregister
constraints = Constraints(angle=(-10, 5), scale=(0.8, 0.2), tx=(60, 3), ty=(20, 1))
imreg_result = similarity(
    master, slave, constraints=constraints, numiter=5, filter_pcorr=5
)

# Transform the slave image
slave_transformed = transform_img(slave, imreg_result.transformation, invert=False)

_, axs = plt.subplots(1, 4, figsize=(13, 8))
im_0 = axs[0].imshow(master)
plt.colorbar(im_0, ax=axs[0])
im_1 = axs[1].imshow(slave)
plt.colorbar(im_1, ax=axs[1])
im_2 = axs[2].imshow(slave_transformed)
plt.colorbar(im_2, ax=axs[2])
im_3 = axs[3].imshow(np.abs(slave_transformed - master))
plt.colorbar(im_3, ax=axs[3])

plt.show()

Features

• Image pre-processing options (frequency filtration, image extension).
• Under-the-hood options exposed (iterations, phase correlation filtration).
• Permissive open-source license (3-clause BSD).
• GPU accelerated

Origin story

This is a fork of the imreg_dft borned of the desire to achieve the following goals:

• Ability to return the final transformation in matrix form as opposed to the angle, translation and scaling factor separately. The original code makes obtaining that matrix really hard because it does it performs using scipy in away that each transformation resizes the image.
• Better performance powered by pytorch
• A more focused codebase. The only goal here is to estimate similarity transformations between pairs of images.

Acknowledgements

The code was originally developed by Christoph Gohlke (University of California, Irvine, USA) and later on developed further by Matěj Týč (Brno University of Technology, CZ). This repo wouldn't exist without them.