def process_images(data_dir, data_labels, size, extension, process=None):
data = []
print('Processing data...')
for f, tags in tqdm(data_labels.values):
fpath = os.path.join(data_dir, '{}.{}'.format(f, extension))
img = io.imread(fpath)
if process == 'RGB':
img = get_rgb(img, [2, 1, 0]) # RGB from TIF
img = np.array(img * 128 + 128, dtype=np.uint8)
elif process == 'NDWI':
img_layer = get_rgb(img, [3, 2, 1]) # NIR-R-G
img = (img_layer[:, :, 2] - img_layer[:, :, 0]) / (
img_layer[:, :, 2] + img_layer[:, :, 0])
img = np.array(img * 128 + 128, dtype=np.uint8)
elif process == 'NDVI':
max_value = 0.7210752894 # evaluated on whole dataset
min_value = -0.71597245876 # evaluated on whole dataset
img = ndvi(img, 2, 3)
img = np.array((img - min_value) * 256 / (max_value - min_value),
dtype=np.uint8)
data.append(cv2.resize(img, (size, size)))
print('Done')
return data
def train_test_images(self, train_positions, test_positions):
img_train, img_test = np.zeros(
shape=(self.height, self.width)), np.zeros(shape=(self.height,
self.width))
for (i, j) in train_positions:
label = self.complete_gt[i, j]
img_train[i, j] = label
for (i, j) in test_positions:
label = self.complete_gt[i, j]
img_test[i, j] = label
return get_rgb(img_train, color_scale=self.color_scale), get_rgb(
img_test, color_scale=self.color_scale)
def tif_to_rgb(img, as_int=True):
# [2, 1, 0] index is needed because colours are stored as bgr (wavelength order)
img_rgb = get_rgb(img, [2, 1, 0])
# scale between 0 and 255 and convert to type uint8
if as_int:
img_rgb = (img_rgb * 255).astype('uint8')
return img_rgb
def display_dataset(img, gt, bands, labels, palette):
"""Display the specified dataset.
Args:
img: 3D hyperspectral image
gt: 2D array labels
bands: tuple of RGB bands to select
labels: list of label class names
palette: dict of colors
display (optional): type of display, if any
"""
print("Image has dimensions {}x{} and {} channels".format(*img.shape))
rgb = spectral.get_rgb(img, bands)
rgb /= np.max(rgb)
rgb = np.asarray(255 * rgb, dtype='uint8')
# Display the RGB composite image
caption = "RGB (bands {}, {}, {})".format(*bands)
# send to visdom server
plt.figure()
# plt.imshow(stddev_stretch(rgb))
plt.imshow(rgb)
plt.xticks(np.array([]))
plt.yticks(np.array([]))
plt.title(caption)
def show_tiff_image_data(bgrn_image):
"""Show a rendering of scaled RGB of BGRN 4 channel image matrix"""
tif_rgb = get_rgb(bgrn_image, [2, 1, 0]) # RGB
# rescaling to 0-255 range - uint8 for display
rescaleIMG = np.reshape(tif_rgb, (-1, 1))
scaler = MinMaxScaler(feature_range=(0, 255))
rescaleIMG = scaler.fit_transform(rescaleIMG)
img_scaled = (np.reshape(rescaleIMG, tif_rgb.shape)).astype(np.uint8)
new_style = {'grid': False}
plt.imshow(img_scaled)
plt.title('RGB')
plt.colorbar()
def preprocess_image1(img, target_size):
if target_size:
img = resize(img, (img.shape[0], target_size[1], target_size[2]),
order=1)
img_rgb = get_rgb(img.transpose(1, 2, 0), [2, 1, 0]) # R-G-B
rescaleimg = np.reshape(img_rgb, (-1, 1))
scaler = MinMaxScaler(feature_range=(0, 255))
rescaleimg = scaler.fit_transform(rescaleimg) # .astype(np.float32)
img_scaled = (np.reshape(rescaleimg, img_rgb.shape)) / 255.0
img_scaled = img_scaled.transpose(2, 0, 1)
img_nir = get_rgb(img.transpose(1, 2, 0), [3, 2, 1]) # NIR-R-G
img_nir_red = (img_nir[:, :, 0] - img_nir[:, :, 1]) / (
img_nir[:, :, 0] + img_nir[:, :, 1] + np.finfo(float).eps
) # (NIR - RED) / (NIR + RED)
img_nir_red = np.expand_dims(np.clip(img_nir_red, -1, 1), axis=0)
img_nir_green = (img_nir[:, :, 2] - img_nir[:, :, 0]) / (
img_nir[:, :, 2] + img_nir[:, :, 0] + np.finfo(float).eps
) # (GREEN - NIR) / (GREEN + NIR)
img_nir_green = np.expand_dims(np.clip(img_nir_green, -1, 1), axis=0)
return np.concatenate((img_scaled, img_nir_red, img_nir_green), axis=0)
def set_data(self, data, bands=None, **kwargs):
'''Sets the data to be shown in the RGB channels.
Arguments:
`data` (ndarray or SpyImage):
If `data` has more than 3 bands, the `bands` argument can be
used to specify which 3 bands to display. `data` will be
passed to `get_rgb` prior to display.
`bands` (3-tuple of int):
Indices of the 3 bands to display from `data`.
Keyword Arguments:
Any valid keyword for `get_rgb` or `matplotlib.imshow` can be
given.
'''
import spectral as spy
rgb_kwargs = {}
for k in ('stretch', 'stretch_all', 'bounds'):
if k in kwargs:
rgb_kwargs[k] = kwargs.pop(k)
# If it is a gray-scale image, only keep the first RGB component so
# matplotlib imshow's cmap can still be used.
self.data = spy.get_rgb(data, bands, **rgb_kwargs)
if len(data.shape) == 2 or (bands is not None and len(bands) == 1):
self.data = self.data[:, :, 0]
if self._image_shape is None:
self._image_shape = data.shape[:2]
elif data.shape[:2] != self._image_shape:
raise ValueError('Image shape is inconsistent with previously ' \
'set data.')
self.imshow_data_kwargs.update(kwargs)
if 'interpolation' in self.imshow_data_kwargs:
self.interpolation = self.imshow_data_kwargs['interpolation']
self.imshow_data_kwargs.pop('interpolation')
if len(kwargs) > 0 and self.is_shown:
msg = 'Keyword args to set_data only have an effect if ' \
'given before the image is shown.'
warnings.warn(UserWarning(msg))
if self.is_shown:
self.refresh()
def convert_tif_pkl(tif_folder='..' + paths.DATA_FOLDER + '/train-tif-sample',
jpg_folder='..' + paths.DATA_FOLDER + '/train-jpg-sample'):
img = np.zeros((256, 256, 3))
scaler = MinMaxScaler(feature_range=(0, 255))
for f in tqdm(os.listdir(tif_folder)):
tif_img = sio.imread(tif_folder + '/' + f)
fil_no = os.path.splitext(f)[0]
jpg_img = sio.imread(jpg_folder + '/' + fil_no + '.jpg')
img2 = get_rgb(tif_img, [2, 1, 0])
rescaleIMG = np.reshape(img2, (-1, 1))
rescaleIMG = scaler.fit_transform(rescaleIMG)
img2 = (np.reshape(rescaleIMG, img2.shape)).astype(np.uint8)
img = img2
print img
plt.imshow(img)
plt.show(img)
def loadImg(self, filePath: str):
if self.imgItem is not None:
self.graphicsView.scene().removeItem(self.imgItem)
self.resultLabel.clear()
self.resultLabel.setText("Click on img")
self.prevPoint = None
file: BilFile = open_image(filePath)
rgb = get_rgb(file)
rgb = rgb * 255
rgb = rgb.astype(np.uint8)
loadedImg = QImage(rgb.tobytes(), rgb.shape[0], rgb.shape[1],
rgb.shape[0] * 3, QImage.Format_RGB888)
self.hyperspectralImg = HyperspectralImgModel(file, loadedImg)
self.__raiseOnImgLoaded(self.hyperspectralImg)
p = QPixmap.fromImage(loadedImg, Qt.AutoColor)
self.imgItem: QGraphicsPixmapItem = self.graphicsView.scene(
).addPixmap(p)
self.scaleImg(self.zoomSlider.value())
def display_dataset(img, gt, bands, labels, palette, vis):
"""Display the specified dataset.
Args:
img: 3D hyperspectral image
gt: 2D array labels
bands: tuple of RGB bands to select
labels: list of label class names
palette: dict of colors
display (optional): type of display, if any
"""
print("Image has dimensions {}x{} and {} channels".format(*img.shape))
rgb = spectral.get_rgb(img, bands)
rgb /= np.max(rgb)
rgb = np.asarray(255 * rgb, dtype='uint8')
# Display the RGB composite image
caption = "RGB (bands {}, {}, {})".format(*bands)
# send to visdom server
vis.images([np.transpose(rgb, (2, 0, 1))], opts={'caption': caption})
def disconnect(self):
self.lasso.disconnect_events()
self.fc[:, -1] = 1
self.collection.set_facecolors(self.fc)
self.canvas.draw_idle()
if __name__ == '__main__':
PROJ_DIR = os.getcwd()
SAMPLES_DIR = os.path.join(PROJ_DIR, 'samples')
sample = SAMPLES_DIR + '\\HARC000.bil.hdr'
s = open_image(sample) #.load()
bands = (70, 70, 30)
rgb = get_rgb(s, bands)
fig, ax = plt.subplots()
im = ax.imshow(rgb)
a = s.asarray()
pts = ax.scatter(a[:, 0], a[:, 1], s=80)
print('Offsets: ', pts.get_offsets())
selector = SelectFromImage(ax, pts)
plt.draw()
def accept(event):
if event.key == "enter":
print("Selected points:")
print(selector.xys[selector.ind])
selector.disconnect()
def output_image(input, output):
return get_rgb(output, color_scale=input.color_scale)
def rgb(img, bands=None):
"""Returns the RGB image from the selected bands (R, G, B)"""
return spectral.get_rgb(img, bands=bands)
def compute_ndwi(bgrn_image):
nrg = get_rgb(bgrn_image, [3, 2, 1]) # NIR-R-G
ndwi = (nrg[:, :, 2] - nrg[:, :, 0]) / \
(nrg[:, :, 2] + nrg[:, :, 0]).astype(np.float64)
return ndwi
def rgb(img, bands=None):
"""Returns the RGB image from the selected bands (R, G, B)"""
from spectral import get_rgb
return get_rgb(img, bands=bands)
from matplotlib.path import Path from spectral import open_image, imshow, get_rgb import matplotlib.pyplot as plt import matplotlib.image as img import matplotlib.cm as cm from spectral import * PROJ_DIR = os.getcwd() SAMPLES_DIR = os.path.join(PROJ_DIR, 'samples') sample = 'D:\\Hypercubes\\HARC000.bil.hdr' image = open_image(sample) #.load() bands = (70, 70, 30) fig, ax = plt.subplots() rgb = get_rgb(image, bands) im = ax.imshow(rgb) plt.draw() plt.show() pdb.set_trace() print(image) x = input() # im = imshow(s.load(), bands) # data = s.read_bands(bands) # im = imshow(s.load(), bands) # plt.show() # plt.imshow(s.read_band(0))
def __call__(self, sample):
image = sample['image']
image = get_rgb(image, bands=[55, 41, 12])
return {'image': image, 'annotation': sample['annotation'].copy()}
