pynterpol

Fast image interpolation in Python.

Installation

python -m pip install pynterpol

Usage (example)

from pynterpol import interp_bilinear_u8
import cv2

img = cv2.imread('image.jpg')  # or any other way of reading an image as a numpy ndarray

interpolated_values = interp_bilinear_u8(img, [100.1, 100.8], [50.5, 51.2])

Performances

High performance is achieved:

by considering that the image to interpolate is defined on an orthogonal equally-spaced regular grid, and
by using Numba for the loop-intensive parts of the computation, and
by optimizing the algorithm implementation (avoid code cache misses for special cases, etc.)

Comparative Timings

Timings realized on an Apple MacBook Pro (Intel Core i9 8 cores, 2.3 GHz) with an gray (i.e. single channel) image of size 1904x1081x1 interpolated on a regular equally-spaced grid of size 3807x2161 that covers the entire domain of the image (that corresponds to a x2 upsampling of the input image).

method	timings
SciPy `RegularGridInterpolator`	1614 ms
`interp_bilinear_u8` (without Numba)	82980 ms (LOL)
`interp_bilinear_u8` (with Numba)	65 ms (almost x25 speedup)

Reference

As of 2022-03-06, only images represented as a NumPy ndarray with uint8 as dtype are supported. Interpolated values are stored as float64. All of this is implemented with a single function: interp_bilinear_u8.

Other cases may be considered in the future:

bicubic interpolation or other interpolation schemes
image represented with different types (float, bigger uints, ...)
integer (or smaller floats) interpolated values

Tests

cd $PROJECT_ROOT
PYTHONPATH=./src python -m unittest tests.test_pynterpol.TestPynterpol

Note: There may be a better way than providing the PYTHONPATH variable in the command...

Distribution

python -m pip install --upgrade build twine
python -m pip build
python -m twine upload --repository testpypi dist/*

FlorentBrunet/pynterpol