def initialize_image(self):
# Creates a local composite of the original image data for display
if self.parent.data.im_type == 'wv02_ms':
self.display_image = utils.create_composite([
self.parent.data.original_image[4, :, :],
self.parent.data.original_image[2, :, :],
self.parent.data.original_image[1, :, :]
],
dtype=np.uint8)
elif self.parent.data.im_type == 'pan':
self.display_image = utils.create_composite([
self.parent.data.original_image,
self.parent.data.original_image,
self.parent.data.original_image
],
dtype=np.uint8)
elif self.parent.data.im_type == 'srgb':
self.display_image = utils.create_composite([
self.parent.data.original_image[0, :, :],
self.parent.data.original_image[1, :, :],
self.parent.data.original_image[2, :, :]
],
dtype=np.uint8)
self.disp_xdim, self.disp_ydim = np.shape(self.display_image)[0:2]
def save_color_image(image_data, output_name, image_type, block_cols,
block_rows):
"""
Write a rgb color image (as png) of the raw image data to disk.
"""
holder = []
# Find the appropriate bands to use for an rgb representation
if image_type == 'wv02_ms':
rgb = [5, 3, 2]
elif image_type == 'srgb':
rgb = [1, 2, 3]
else:
rgb = [1, 1, 1]
red_band = image_data[rgb[0]]
green_band = image_data[rgb[1]]
blue_band = image_data[rgb[2]]
for i in range(len(red_band)):
holder.append(
utils.create_composite([red_band[i], green_band[i], blue_band[i]]))
colorfullimg = utils.compile_subimages(holder, block_cols, block_rows, 3)
mimg.imsave(output_name, colorfullimg)
colorfullimg = None
def display_watershed(original_data, watershed_data, block=5):
# block = 5
watershed = watershed_data[block]
original_1 = original_data[6][block]
original_2 = original_data[4][block]
original_3 = original_data[1][block]
# randcolor = colors.ListedColormap(np.random.rand(256,3))
ws_bound = segmentation.find_boundaries(watershed)
ws_display = utils.create_composite([original_1, original_2, original_3])
ws_display[:, :, 0][ws_bound] = 240
ws_display[:, :, 1][ws_bound] = 80
ws_display[:, :, 2][ws_bound] = 80
display_im = utils.create_composite([original_1, original_2, original_3])
fig, axes = plt.subplots(1, 2, subplot_kw={'xticks': [], 'yticks': []})
fig.subplots_adjust(hspace=0.3, wspace=0.05)
# axes[1].imshow(self.sobel_image,interpolation='none',cmap='gray')
axes[0].imshow(display_im, interpolation='none')
axes[1].imshow(ws_display, interpolation='none')
plt.show()
def watershed_transformation(image_data, sobel_threshold,
amplification_factor):
'''
Runs a watershed transform on the main dataset
1. Create a gradient image using the sobel algorithm
2. Adjust the gradient image based on given threshold and amplification.
3. Find the local minimum gradient values and place a marker
4. Construct watersheds on top of the gradient image starting at the
markers.
5. Recombine neighboring image segments using a region adjacency graph.
'''
# If this block has no data, return a placeholder watershed.
if np.amax(image_data[0]) <= 1:
return np.zeros(np.shape(image_data))
# Create a gradient image using a sobel filter
sobel_image = filters.sobel(image_data[2])
# Adjust the sobel image based on the given threshold and amp factor.
upper_threshold = np.amax(sobel_image) / amplification_factor
if upper_threshold < 0.20:
upper_threshold = 0.20
sobel_image = exposure.rescale_intensity(sobel_image,
in_range=(0, upper_threshold),
out_range=(0, 1))
# Prevent the watersheds from 'leaking' along the sides of the image
sobel_image[:, 0] = 1
sobel_image[:, -1] = 1
sobel_image[0, :] = 1
sobel_image[-1, :] = 1
# Set all values in the sobel image that are lower than the
# given threshold to zero.
sobel_image[sobel_image <= sobel_threshold] = 0
#sobel_copy = np.copy(sobel_image)
# Find local minimum values in the sobel image by inverting
# sobel_image and finding the local maximum values
inv_sobel = 1 - sobel_image
local_min = feature.peak_local_max(inv_sobel,
min_distance=2,
indices=False,
num_peaks_per_label=1)
markers = ndimage.label(local_min)[0]
# Build a watershed from the markers on top of the edge image
im_watersheds = morphology.watershed(sobel_image, markers)
im_watersheds = np.array(im_watersheds, dtype='uint16')
# Clear gradient image data
sobel_image = None
# Find the locations that contain no spectral data
empty_pixels = np.zeros(np.shape(image_data[0]), dtype='bool')
empty_pixels[image_data[2] == 0] = True
# Set all values outside of the image area (empty pixels, usually caused by
# orthorectification) to one value, at the end of the watershed list.
im_watersheds[empty_pixels] = np.amax(im_watersheds) + 1
# Recombine segments that are adjacent and similar to each other.
# Created a region adjacency graph. Create_composite() takes a single list
# of bands
color_im = utils.create_composite(
[image_data[0], image_data[1], image_data[2]])
# Clear image data
image_data = None
# Create the region adjacency graph based on the color image
try:
im_graph = graph.rag_mean_color(color_im, im_watersheds)
except KeyError:
pass
else:
# Clear color image data
color_im = None
# Combine segments that are adjacent and whose pixel intensity
# difference is less than 10.
im_watersheds = graph.cut_threshold(im_watersheds, im_graph, 5.0)
return im_watersheds
def mode_one(segment_list, label_vector, feature_matrix, input_directory, im_type, tds_filename):
## Build a list of the images in the input directory
# for each image in the directory, image list contains the "..._segmented.h5" file if it exists,
# and the raw image file if it does not.
# If there is an existing segment list, note the required image to continue the
# training set
if segment_list:
required_images = get_required_images(segment_list[len(label_vector):])
else:
required_images = []
# Add the images in the provided folder to the image list
seg_list = list(utils.get_image_paths(input_directory, keyword='segmented.h5', strict=False))
for ext in utils.valid_extensions:
raw_list = utils.get_image_paths(input_directory, keyword=ext)
for raw_im in raw_list:
if os.path.splitext(raw_im)[0] + "_segmented.h5" not in seg_list:
seg_list.append(raw_im)
# Save only the unique entries
image_list = list(set(seg_list))
utils.remove_hidden(image_list)
# Make sure we have all of the required images
for image in required_images:
if image in image_list:
print "Missing required image: {}".format(image)
quit()
# **************************************** #
# For expanding icebridge training set 9.18.18
target_images = ["2016_07_19_01052.JPG",
"2017_07_25_01613.JPG",
"2017_07_25_01772.JPG",
"2017_07_25_01783.JPG",
"2017_07_25_06767.JPG"]
# **************************************** #
# As long as there were files in the input directory, loop indefinitely
# This loop breaks when the training GUI is closed.
while image_list:
# Cycle through each image in the list
for next_image in image_list:
# If we are out of predetermined segments to classify, start picking
# images from those in the directory provided
if len(segment_list) == len(label_vector):
image_name = os.path.split(next_image)[1]
# if image_name not in target_images:
# print("Skipping {}".format(image_name))
# continue
# Otherwise find the image name from the next unlabeled segment
else:
image_name = segment_list[len(label_vector)][0]
# Convert the image name into the segmented name. This file must
# exist to work from an existing tds. That is, we cant resegment the
# image, because it will give different results and segment ids will not match.
image_name = os.path.splitext(image_name)[0] + "_segmented.h5"
# image_root, image_ext = os.path.splitext(image_name)
print("Working on image: {}".format(image_name))
# If the image is already segmented
if image_name.split('_')[-1] == 'segmented.h5':
segmented_name = image_name
# Otherwise image_name should point to a raw image
else:
print("Creating segments on provided image...")
# Preprocess next_image data
image_data, meta_data = prepare_image(input_directory, image_name, im_type,
output_path=input_directory, verbose=True)
# Create image segments
segmented_name = os.path.splitext(image_name)[0] + '_segmented.h5'
seg_path = os.path.join(input_directory, segmented_name)
segment_image(image_data, image_type=im_type, write_results=True,
dst_file=seg_path, verbose=True)
# Convert this entry in image_list to the "..._segmented.h5" name
image_list.append(segmented_name)
image_list.remove(next_image)
h5_file = os.path.join(input_directory, segmented_name)
# from segment import load_from_disk
original_image = []
original_image_dict, im_type = load_from_disk(h5_file, False)
for sub_image in range(len(original_image_dict[1])):
# Create a list of image blocks based on the number of bands
# in the input image.
# Need to verify this for pan images (ncw 9.17.18):
if im_type == 'pan':
original_image.append(original_image_dict[1])
if im_type == 'wv02_ms':
original_image.append(
utils.create_composite([original_image_dict[i][sub_image] for i in range(1,9)],
dtype=c_int)
)
if im_type == 'srgb':
original_image.append(
utils.create_composite([original_image_dict[i][sub_image] for i in range(1,4)],
dtype=c_int)
)
with h5py.File(h5_file, 'r') as f:
im_date = f.attrs.get("Image Date")
watershed_image = f['watershed'][:]
# Convert the segmented image to c_int datatype. This is needed for the
# Cython methods that calculate attribute of segments.
watershed_image = np.ndarray.astype(watershed_image, c_int)
#### Initializing the GUI
tW = TrainingWindow(original_image, watershed_image, segment_list,
image_name, label_vector, feature_matrix, im_type, im_date,
1, tds_filename)
image_list.remove(next_image)
h5_file = os.path.join(input_path, segmented_name)
# from segment import load_from_disk
original_image = []
original_image_dict, im_type = load_from_disk(h5_file, False)
for sub_image in range(len(original_image_dict[1])):
# Create a list of image blocks based on the number of bands
# in the input image.
if im_type == 'wv02_ms':
original_image.append(
utils.create_composite([
original_image_dict[1][sub_image],
original_image_dict[2][sub_image],
original_image_dict[3][sub_image],
original_image_dict[4][sub_image],
original_image_dict[5][sub_image],
original_image_dict[6][sub_image],
original_image_dict[7][sub_image],
original_image_dict[8][sub_image]
]))
if im_type == 'srgb':
original_image.append(
utils.create_composite([
original_image_dict[1][sub_image],
original_image_dict[2][sub_image],
original_image_dict[3][sub_image]
]))
# original_image = utils.create_composite([original_image[1],
# original_image[2],
# original_image[3]])
def display_image(raw, watershed, classified, type):
# Save a color
empty_color = [.1, .1, .1] #Almost black
snow_color = [.9, .9, .9] #Almost white
pond_color = [.31, .431, .647] #Blue
gray_color = [.65, .65, .65] #Gray
water_color = [0., 0., 0.] #Black
shadow_color = [.100, .545, .0] #Orange
custom_colormap = [
empty_color, snow_color, gray_color, pond_color, water_color,
shadow_color
]
custom_colormap = colors.ListedColormap(custom_colormap)
#Making sure there is atleast one of every pixel so the colors map properly (only changes
# display image, not saved data)
classified[0][0] = 0
classified[1][0] = 1
classified[2][0] = 2
classified[3][0] = 3
classified[4][0] = 4
classified[5][0] = 5
# Figure that show 3 images: raw, segmented, and classified
if type == 1:
# Creating the watershed display image with borders highlighted
ws_bound = segmentation.find_boundaries(watershed)
ws_display = utils.create_composite([raw, raw, raw])
ws_display[:, :, 0][ws_bound] = 255
ws_display[:, :, 1][ws_bound] = 255
ws_display[:, :, 2][ws_bound] = 22
fig, axes = plt.subplots(1, 3, subplot_kw={'xticks': [], 'yticks': []})
fig.subplots_adjust(left=0.05,
right=0.99,
bottom=0.05,
top=0.90,
wspace=0.02,
hspace=0.2)
tnrfont = {'fontname': 'Times New Roman'}
axes[0].imshow(raw, cmap='gray', interpolation='None')
axes[0].set_title("Raw Image", **tnrfont)
axes[1].imshow(ws_display, interpolation='None')
axes[1].set_title("Image Segments", **tnrfont)
axes[2].imshow(classified, cmap=custom_colormap, interpolation='None')
axes[2].set_title("Classification Output", **tnrfont)
# Figure that shows 2 images: raw and classified.
if type == 2:
fig, axes = plt.subplots(1, 2, subplot_kw={'xticks': [], 'yticks': []})
fig.subplots_adjust(hspace=0.3, wspace=0.05)
axes[0].imshow(raw, interpolation='None')
axes[0].set_title("Raw Image")
axes[1].imshow(classified, cmap=custom_colormap, interpolation='None')
axes[1].set_title("Classification Output")
plt.show()
def classify_image(input_image, watershed_data, training_dataset, meta_data,
threads=1, quality_control=False, verbose=False):
'''
Run a random forest classification.
Input:
input_image: preprocessed image data (preprocess.py)
watershed_image: Image objects created with the segmentation
algorithm. (segment.py)
training_dataset: Tuple of training data in the form:
(label_vector, attribute_matrix)
meta_data: [im_type, im_date]
Returns:
Raster of classified data.
'''
#### Prepare Data and Variables
num_blocks = len(input_image[1])
num_bands = len(input_image.keys())
image_type = meta_data[0]
image_date = meta_data[1]
## Restructure the input data.
# We are creating a single list where each element of the list is one
# block (old: subimage) of the image and is a stack of all bands.
image_data = [] # [block:row:column:band]
for blk in range(num_blocks):
image_data.append(utils.create_composite(
[input_image[b][blk] for b in range(1,num_bands+1)]))
input_image = None
## Parse training_dataset input
label_vector = training_dataset[0]
training_feature_matrix = training_dataset[1]
# Method for assessing the quality of the training dataset.
# quality_control = True
# if quality_control == True:
# debug_tools.test_training(label_vector, training_feature_matrix)
# aa = raw_input("Continue? ")
# if aa == 'n':
# quit()
#### Construct the random forest decision tree using the training data set
rfc = RandomForestClassifier(n_estimators=100)
rfc.fit(training_feature_matrix, label_vector)
#### Classify each image block
# Define multiprocessing-safe queues containing data to process
clsf_block_queue = Queue()
num_blocks = len(watershed_data)
im_block_queue = construct_block_queue(image_data, watershed_data, num_blocks)
# Define the number of threads to create
NUMBER_OF_PROCESSES = threads
block_procs = [Process(target=process_block_helper,
args=(im_block_queue, clsf_block_queue, image_type,
image_date, rfc))
for _ in range(NUMBER_OF_PROCESSES)]
# Start the worker processes.
for proc in block_procs:
# Add a stop command to the end of the queue for each of the
# processes started. This will signal for the process to stop.
im_block_queue.put('STOP')
# Start the process
proc.start()
# Display a progress bar
if verbose:
try:
from tqdm import tqdm
except ImportError:
print "Install tqdm to display progress bar."
verbose = False
else:
pbar = tqdm(total=num_blocks, unit='block')
# Each process adds the classification results to clsf_block_queue, when it
# finishes a row. Adds 'None' when there are not more rows left
# in the queue.
# This loop continues as long as all of the processes have not finished
# (i.e. fewer than NUMBER_OF_PROCESSES have returned None). When a row is
# added to the queue, the tqdm progress bar updates.
# Initialize the output dataset as an empty list of length = input dataset
# This needs to be initialized since blocks will be added non-sequentially
clsf_block_list = [None for _ in range(num_blocks)]
finished_threads = 0
while finished_threads < NUMBER_OF_PROCESSES:
if not clsf_block_queue.empty():
val = clsf_block_queue.get()
if val == None:
finished_threads += 1
else:
block_num = val[0]
segmnt_data = val[1]
clsf_block_list[block_num] = segmnt_data
if verbose: pbar.update()
# Close the progress bar
if verbose:
pbar.close()
print "Finished Processing. Closing threads..."
# Join all of the processes back together
for proc in block_procs:
proc.join()
return clsf_block_list
def classify_image(input_image,
watershed_data,
training_dataset,
meta_data,
threads=2,
quality_control=False,
debug_flag=False,
verbose=False):
'''
Run a random forest classification.
Input:
input_image: preprocessed image data (preprocess.py)
watershed_image: Image objects created with the segmentation
algorithm. (segment.py)
training_dataset: Tuple of training data in the form:
(label_vector, attribute_matrix)
meta_data: [im_type, im_date]
Returns:
Raster of classified data.
'''
#### Prepare Data and Variables
num_blocks = len(input_image[1])
num_bands = len(input_image.keys())
image_type = meta_data[0]
image_date = meta_data[1]
## Restructure the input data.
# We are creating a single list where each element of the list is one
# block (old: subimage) of the image and is a stack of all bands.
image_data = [] # [block:row:column:band]
for blk in range(num_blocks):
image_data.append(
utils.create_composite(
[input_image[b][blk] for b in range(1, num_bands + 1)]))
input_image = None
# watershed_image = input_file['watershed'][:]
# watershed_dimensions = input_file['dimensions'][:]
# num_x_subimages = dimensions[0]
# num_y_subimages = dimensions[1]
## Parse training_dataset input
label_vector = training_dataset[0]
training_feature_matrix = training_dataset[1]
# im_type = input_file.attrs.get('Image Type')
# im_date = input_file.attrs.get('Image Date')
#Method for assessing the quality of the training dataset.
if quality_control == True:
test_training(label_vector, training_feature_matrix)
aa = raw_input("Continue? ")
if aa == 'n':
quit()
# # If there is no information in this image file, save a dummy classified image and exit
# # This can often happen depending on the original image dimensions and the amount it was split
# if np.sum(band_1) == 0:
# classified_image_path = os.path.join(output_filepath, output_filename + '_classified_image.png')
# outfile = h5py.File(os.path.join(output_filepath, output_filename + '_classified.h5'),'w')
# if im_type == 'wv02_ms':
# empty_bands = np.zeros(np.shape(band_1)[0],np.shape(band_1)[1],8)
# empty_image = utils.compile_subimages(empty_bands, num_x_subimages, num_y_subimages, 8)
# elif im_type == 'srgb':
# empty_bands = np.zeros(np.shape(band_1)[0],np.shape(band_1)[1],3)
# empty_image = utils.compile_subimages(empty_bands, num_x_subimages, num_y_subimages, 3)
# elif im_type == 'pan':
# empty_image = np.zeros(np.shape(band_1))
# outfile.create_dataset('classified', data=empty_image,compression='gzip',compression_opts=9)
# outfile.create_dataset('original', data=empty_image,compression='gzip',compression_opts=9)
# outfile.close()
# # return a 1x5 array with values of one for the pixel counts
# return output_filename, np.ones(5)
#### Construct the random forest decision tree using the training data set
rfc = RandomForestClassifier()
rfc.fit(training_feature_matrix, label_vector)
#### Classify each image block
# Define multiprocessing-safe queues containing data to process
clsf_block_queue = Queue()
num_blocks = len(watershed_data)
im_block_queue = construct_block_queue(image_data, watershed_data,
num_blocks)
# Define the number of threads to create
NUMBER_OF_PROCESSES = threads
block_procs = [
Process(target=process_block_helper,
args=(im_block_queue, clsf_block_queue, image_type, image_date,
rfc)) for _ in range(NUMBER_OF_PROCESSES)
]
# Start the worker processes.
for proc in block_procs:
# Add a stop command to the end of the queue for each of the
# processes started. This will signal for the process to stop.
im_block_queue.put('STOP')
# Start the process
proc.start()
# Display a progress bar
if verbose:
try:
from tqdm import tqdm
except ImportError:
print "Install tqdm to display progress bar."
verbose = False
else:
pbar = tqdm(total=num_blocks, unit='block')
# Each process adds the classification results to clsf_block_queue, when it
# finishes a row. Adds 'None' when there are not more rows left
# in the queue.
# This loop continues as long as all of the processes have not finished
# (i.e. fewer than NUMBER_OF_PROCESSES have returned None). When a row is
# added to the queue, the tqdm progress bar updates.
# Initialize the output dataset as an empty list of length = input dataset
# This needs to be initialized since blocks will be added non-sequentially
clsf_block_list = [None for _ in range(num_blocks)]
finished_threads = 0
while finished_threads < NUMBER_OF_PROCESSES:
if not clsf_block_queue.empty():
val = clsf_block_queue.get()
if val == None:
finished_threads += 1
else:
block_num = val[0]
segmnt_data = val[1]
clsf_block_list[block_num] = segmnt_data
if verbose: pbar.update()
# Close the progress bar
if verbose:
pbar.close()
print "Finished Processing. Closing threads..."
# Join all of the processes back together
for proc in block_procs:
proc.join()
return clsf_block_list
# Lite version: Save only the classified output, and do not save the original image data
compiled_classified = utils.compile_subimages(classified_image,
num_x_subimages,
num_y_subimages, 1)
if verbose: print "Saving..."
with h5py.File(
os.path.join(output_filepath, output_filename + '_classified.h5'),
'w') as outfile:
outfile.create_dataset('classified',
data=compiled_classified,
compression='gzip',
compression_opts=9)
#### Count the number of pixels that were in each classification category.
sum_snow, sum_gray_ice, sum_melt_ponds, sum_open_water, sum_shadow = utils.count_features(
compiled_classified)
pixel_counts = [
sum_snow, sum_gray_ice, sum_melt_ponds, sum_open_water, sum_shadow
]
# Clear the image datasets from memory
compiled_classified = None
input_image = None
watershed_image = None
cur_image = None
cur_ws = None
entropy_image = None
if verbose: print "Done."
return output_filename, pixel_counts
