Automated processing of historical aerial photography. Quickly crop images, match histograms, detect fiducials, and mask labels and borders.
import cv2
import matplotlib.pyplot as plt
import numpy as np
import os
from aerio.Photo import Photo
from aerio.PhotoCollection import PhotoCollection
from aerio.BoundingBoxCollection import BoundingBoxCollection
from aerio import utilsBefore processing, photos must be loaded into a PhotoCollection
# Select a directory that contains multiple photos
photo_dir = os.path.join("data")
# Generate absolute paths to each photo
photo_paths = utils.list_paths(photo_dir, valid_extensions=[".tif"])
# Load all photos in the directory as a Photo Collection
collection = PhotoCollection(photo_paths, photo_size=(224, 224))Photos within a PhotoCollection can be described and previewed
# Use indexes to select single photos from within a collection
photo = collection[0]
# Printing a photo displays various parameters
print(photo)
# Previewing a photo displays a thumbnail for visual reference
photo.preview(size=(8, 8))DWW_3FF_117
Resolution (px): 1886 (H) x 1885 (W)
DPI: 213.8022
Size (mm): 224 (H) x 224 (W)
Pixel size (mm): 0.1188 (H) x 0.1188 (W)
Pre-processing can be used to make sure all photos within a collection are equal size and radiometrically normalized, which can improve aerial triangulation accuracy.
# Crop all photos to the minimum photo dimensions
collection.crop()
# Match histograms for all photos using one of the photos as a reference
collection.match_histograms()# After processing, photo images can be saved to disk for later use
collection.save(path=os.path.join("data", "processed"), suffix="preprocessed")
# Photo images can also be extracted as a list of arrays for use outside of aerio
collection.images;Fiducial markers may need to be located to perform internal alignment between photos. Currently, aerio only supports automatic location of notched fiducials, as shown in this example.
# Automatically locate the fiducial point in each photo. The size argument indicates
# the approximate height and width of the fiducial. If fiducials cannot be accurately
# located, try adjusting the size to include less or more of the fiducial.
collection.locate_fiducials(size=(80, 120))
# Individual fiducials can be previewed to confirm location accuracy
collection[0].fiducials.bottom.preview()
# Coordinates can also be returned as a list of tuples
collection[0].fiducials.coordinates[(939.0, 54.0), (1832.0, 946.0), (938.0, 1833.0), (53.0, 945.0)]
Historical aerial photos often contain elements such as borders and labels that are not part of the image and may interfere with aerial triangulation. aerio contains tools to mask these objects within individual photos using bounding boxes.
To mask labels, their bounding boxes must first be located. This can be done manually by entering a list of pixel coordinates. Coordinates can be entered clockwise or counterclockwise. If labels are at the same position in every image, manually locating labels may be worth the hassle.
photo = collection[0]
# Boxes can be manually created as a list of pixel coordinates. Coordinates
# can be entered clockwise or counterclockwise.
upper_left = [[50, 50], [350, 50], [350, 120], [50, 120]]
upper_right = [[1400, 50], [1830, 50], [1830, 120], [1400, 120]]
labels = BoundingBoxCollection([upper_left, upper_right], photo)
labels.preview()
There are many techniques for automatically locating text within images. In this example, we will use the swtloc library to generate bounding boxes around text. This method can handle labels that are in different positions within different images, but is prone to locating false positive labels and typically requires parameter tuning and post-processing to achieve good results.
from swtloc import SWTLocalizer
swtl = SWTLocalizer()
swtl.swttransform(image=photo.img, save_results=False,
edge_func = 'ac', ac_sigma = 0.4, text_mode = 'db_lf',
gs_blurr=True, blurr_kernel = (15, 15),
minrsw = 5, maxrsw = 15,
minCC_comppx = 50, maxCC_comppx = 1000,
max_angledev = np.pi/6,
acceptCC_aspectratio = 5.)
bbox, _ = swtl.get_extreme_bbox(show=False)
# Turn the lists of coordinates into Bounding Box objects
labels = BoundingBoxCollection(bbox, photo)
# Preview the un-processed bounding boxes. Note the large number of false
# positive boxes.
labels.preview(color=(0, 255, 0))Transforming...
Operation Completed!
To remove false positive labels, aerio has tools for cleaning up bounding boxes.
# Collapse nearby bounding boxes to merge characters. An asymetrical kernel
# helps to connect horizontally adjacent characters.
labels.collapse(kernel=np.ones((5, 15), np.uint8), iterations=20)
# Filter for boxes near the edge of the image that are wider than tall.
labels.filter(max_edge_distance = 100, max_hw_ratio=0.5)
labels.preview(color=(0, 255, 0))
Aerial photo margins often contain borders that can interfere with aerial triangulation. aerio can automatically generate bounding boxes around photo borders. This will only work if the borders are relatively straight and centered.
border = photo.border_box(30)
border.preview()
Additional boxes can be added to a BoundingBoxCollection by adding a list of coordinates or adding the boxes from another collection. This can be used, for example, to combine label boxes and border boxes into a single collection that can be masked.
combined = border + labels
combined.preview()
Bounding boxes can be converted into mask images for use in other software. Images can be saved directly to file or returned as arrays.
# Save the mask to file
combined.save_mask(path=os.path.join("data", "masks"))
# Or return the mask as an array
mask = combined.generate_mask()
# Preview the mask
_, ax = plt.subplots(figsize=(8, 8))
ax.imshow(mask, cmap="gray");