stitch2D is a Python script that stitches a two-dimensional grid of tiles into a mosaic. It was originally developed for stitching together images collected on various microscopes in the Department of Mineral Sciences at the Smithsonian National Museum of Natural History.
When tiles are stitched together by stitch2d, they are translated, not rotated, resized, or warped. As a result, stitch2d requires all images to be the same size and orientation. Images must overlap, although they don’t necessarily need to be arranged in a grid.
In addition to the instructions below, a guide and API reference are available in the documentation.
Install stitch2d with pip:
pip install stitch2d
Or install from the GitHub repository using git and pip:
git clone https://github.com/adamancer/stitch2d
cd stitch2d
pip install .
The following code can be used to create and save a mosaic:
from stitch2d import create_mosaic
mosaic = create_mosaic("/path/to/tiles")
try:
mosaic.load_params()
except FileNotFoundError:
mosaic.downsample(0.6)
mosaic.align()
mosaic.reset_tiles()
mosaic.save_params()
mosaic.smooth_seams()
mosaic.save("mosaic.jpg")
A simple stitching workflow is also available from the command line. To create a smoothed mosaic and save it as a JPEG, run:
stitch2d path/to/tiles --smooth -output mosaic.jpg
For more information about using this command, including available parameters, run:
stitch2d --help
stitch2d includes two classes that can be used to create mosaics from a list of tiles:
Mosaic
, which incorporates no information about how the tiles in
the mosaic are arrangedStructuredMosaic
, which arranges the tiles into a grid based on
parameters supplied by the userYou can also use create_mosaic()
, as above, which accepts the same
arguments as StructuredMosaic
. This function returns a
StructuredMosaic
if grid parameters are provided or can be inferred
from the filenames of the tiles or a Mosaic
if not.
Since Mosaic
doesn’t know anything about the tile structure, it can be
slow, especially for large grids where lots of tiles need to be
compared. It’s almost always faster to use StructuredMosaic
where
possible.
Initialize a Mosaic
by pointing it to the directory where the tiles of
interest live:
from stitch2d import Mosaic
mosaic = Mosaic("/path/to/tiles")
Mosaic
also includes a class attribute, num_cores
, to specify how
many cores it should use when aligning and stitching a mosaic. By
default, it uses one core. Modify this value with:
Mosaic.num_cores = 2
Even when using multiple cores, detecting and extracting features can be time consuming. One way to speed up the process is to reduce the resolution of the tiles being analyzed:
mosaic.downsample(0.6) # downsamples all tiles larger than 0.6 mp
Alternatively you can resize the tiles without the size check:
mosaic.resize(0.6) # resizes all tiles to 0.6 mp
You can then align the smaller tiles:
mosaic.align()
In either case, you can restore the full-size images prior to stitching the mosaic together:
mosaic.reset_tiles()
Sometimes brightness and contrast can vary significantly between
adjacent tiles, producing a checkerboard effect when the mosaic is
stitched together. This can be mitigated in many cases using
smooth_seams()
, which aligns brightness/contrast between neighboring
tiles by comparing areas of overlap:
mosaic.smooth_seams()
Once the tiles have been positioned, the mosaic can be viewed:
mosaic.show()
Or saved to a file:
mosaic.save("mosaic.tif")
Or returned as a numpy array if you need more control over the final mosaic:
arr = mosaic.stitch()
The default backend, opencv, orders color channels as BGR. You may want to reorder the color channels before working with the image in a different program. To get an RGB image from a BGR image, use:
arr = arr[...,::-1].copy()
New in 1.1: Or specify the desired channel order when stitching:
arr = mosaic.stitch("RGB")
Once the tiles are positioned, their locations are stored in the
params
attribute, which can be saved as JSON:
mosaic.save_params("params.json")
Those parameters can then be loaded into a new mosaic if needed:
mosaic.load_params("params.json")
StructuredMosaic
allows the user to specify how the tiles in the
mosaic should be arranged. For tilesets of known structure, it is
generally faster but otherwise works the same as Mosaic
. Initialize a
structured mosaic with:
from stitch2d import StructuredMosaic
mosaic = StructuredMosaic(
"/path/to/tiles",
dim=15, # number of tiles in primary axis
origin="upper left", # position of first tile
direction="horizontal", # primary axis (i.e., the direction to traverse first)
pattern="snake" # snake or raster
)
For large tilesets where adequate-but-imperfect tile placement is
acceptable, StructuredMosaic
can use its knowledge of the tile grid to
quickly build a mosaic based on the positions of only a handful of
tiles:
# Stop aligning once 5 tiles have been successfully placed
mosaic.align(limit=5)
# Build the rest of the mosaic based on the positioned tiles. If from_placed
# is True, missing tiles are appended to the already positioned tiles. If
# False, a new mosaic is calculated from scratch.
mosaic.build_out(from_placed=True)
The build_out()
method can also be used to ensure that all tiles
(including those that could not be placed using feature matching) appear
in the final mosaic. The primary disadvantage of this method is that the
placement of those tiles is less precise.
New in 1.2: The Tile class now includes a prep_imdata()
method
that can be used to tweak the image data being used to align the mosaic.
When using the default OpenCVTile class, this method creates an 8-bit
copy of the image data to use for feature detection and matching while
retaining the original data to use when building the mosaic.
The default behavior of prep_imdata()
is simplistic. To customize it,
use a subclass. For example, the default method scales the intensities
of the original data based on the maximum intensity found in the array.
For images with a small number of extremely bright pixels, this can
yield unusably dim images. A better approach may be to use
np.percentile()
:
import numpy as np
class MyTile(OpenCVTile):
def prep_imdata(self):
imdata = self.imdata - self.imdata.min()
return np.uint8(255 * imdata / np.percentile(imdata, 99))
mosaic = create_mosaic("path/to/tiles", tile_class=MyTile)
The opencv package includes powerful tools for stitching 2D and 3D images. Much of that functionality has been ported to Python as the stitching package, which streamlines the opencv API and includes a useful tutorial. I didn’t have any luck getting it to work consistently with microscope tilesets, but it includes advanced features missing from this package (lens corrections, affine transformations beyond simple translation, etc.) and can be configured to work with 2D images. It’s definitely worth a look for tilesets more complex than the simple case handled here.
Fiji also includes a 2D/3D stitching tool.