def test_invert_bool():
dtype = 'bool'
image = np.zeros((3, 3), dtype=dtype)
image[1, :] = dtype_limits(image)[1]
expected = np.zeros((3, 3), dtype=dtype) + dtype_limits(image)[1]
expected[1, :] = 0
result = invert(image)
assert_array_equal(expected, result)
def test_area_closing(self):
"Test for Area Closing (2 thresholds, all types)"
# original image
img = np.array(
[[240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240],
[240, 200, 200, 240, 200, 240, 200, 200, 240, 240, 200, 240],
[240, 200, 40, 240, 240, 240, 240, 240, 240, 240, 40, 240],
[240, 240, 240, 240, 100, 240, 100, 100, 240, 240, 200, 240],
[240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240],
[200, 200, 200, 200, 200, 200, 200, 240, 200, 200, 255, 255],
[200, 255, 200, 200, 200, 255, 200, 240, 255, 255, 255, 40],
[200, 200, 200, 100, 200, 200, 200, 240, 255, 255, 255, 255],
[200, 200, 200, 100, 200, 200, 200, 240, 200, 200, 255, 255],
[200, 200, 200, 200, 200, 40, 200, 240, 240, 100, 255, 255],
[200, 40, 255, 255, 255, 40, 200, 255, 200, 200, 255, 255],
[200, 200, 200, 200, 200, 200, 200, 255, 255, 255, 255, 255]],
dtype=np.uint8)
# expected area closing with area 2
expected_2 = np.array(
[[240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240],
[240, 200, 200, 240, 240, 240, 200, 200, 240, 240, 200, 240],
[240, 200, 200, 240, 240, 240, 240, 240, 240, 240, 200, 240],
[240, 240, 240, 240, 240, 240, 100, 100, 240, 240, 200, 240],
[240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240],
[200, 200, 200, 200, 200, 200, 200, 240, 200, 200, 255, 255],
[200, 255, 200, 200, 200, 255, 200, 240, 255, 255, 255, 255],
[200, 200, 200, 100, 200, 200, 200, 240, 255, 255, 255, 255],
[200, 200, 200, 100, 200, 200, 200, 240, 200, 200, 255, 255],
[200, 200, 200, 200, 200, 40, 200, 240, 240, 200, 255, 255],
[200, 200, 255, 255, 255, 40, 200, 255, 200, 200, 255, 255],
[200, 200, 200, 200, 200, 200, 200, 255, 255, 255, 255, 255]],
dtype=np.uint8)
# expected diameter closing with diameter 4
expected_4 = np.array(
[[240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240],
[240, 200, 200, 240, 240, 240, 240, 240, 240, 240, 240, 240],
[240, 200, 200, 240, 240, 240, 240, 240, 240, 240, 240, 240],
[240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240],
[240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240, 240],
[200, 200, 200, 200, 200, 200, 200, 240, 240, 240, 255, 255],
[200, 255, 200, 200, 200, 255, 200, 240, 255, 255, 255, 255],
[200, 200, 200, 200, 200, 200, 200, 240, 255, 255, 255, 255],
[200, 200, 200, 200, 200, 200, 200, 240, 200, 200, 255, 255],
[200, 200, 200, 200, 200, 200, 200, 240, 240, 200, 255, 255],
[200, 200, 255, 255, 255, 200, 200, 255, 200, 200, 255, 255],
[200, 200, 200, 200, 200, 200, 200, 255, 255, 255, 255, 255]],
dtype=np.uint8)
# _full_type_test makes a test with many image types.
_full_type_test(img, 2, expected_2, area_closing, connectivity=2)
_full_type_test(img, 4, expected_4, area_closing, connectivity=2)
P, S = max_tree(invert(img), connectivity=2)
_full_type_test(img, 4, expected_4, area_closing,
parent=P, tree_traverser=S)
def test_invert_bool():
dtype = 'bool'
image = np.zeros((3, 3), dtype=dtype)
upper_dtype_limit = dtype_limits(image, clip_negative=False)[1]
image[1, :] = upper_dtype_limit
expected = np.zeros((3, 3), dtype=dtype) + upper_dtype_limit
expected[1, :] = 0
result = invert(image)
assert_array_equal(expected, result)
def test_invert_float64_unsigned():
dtype = 'float64'
image = np.zeros((3, 3), dtype=dtype)
lower_dtype_limit, upper_dtype_limit = \
dtype_limits(image, clip_negative=True)
image[2, :] = upper_dtype_limit
expected = np.zeros((3, 3), dtype=dtype)
expected[0, :] = upper_dtype_limit
expected[1, :] = upper_dtype_limit
result = invert(image)
assert_array_equal(expected, result)
def test_invert_int8():
dtype = 'int8'
image = np.zeros((3, 3), dtype=dtype)
lower_dtype_limit, upper_dtype_limit = \
dtype_limits(image, clip_negative=False)
image[1, :] = lower_dtype_limit
image[2, :] = upper_dtype_limit
expected = np.zeros((3, 3), dtype=dtype)
expected[2, :] = lower_dtype_limit
expected[1, :] = upper_dtype_limit
expected[0, :] = -1
result = invert(image)
assert_array_equal(expected, result)
def segment(self, img, include_intermediate_results=False, **kwargs):
assert img.ndim == 3, 'Expecting 3D image, got shape {}'.format(
img.shape)
img = ndi.median_filter(img, size=(1, 3, 3))
img = img_as_float(img)
img = util.invert(img)
img_mz = img.max(axis=0)
img_mz = exposure.rescale_intensity(img_mz, out_range=(0, 1))
peaks, img_dog, sigmas = blob_dog(img_mz,
min_sigma=8,
max_sigma=128,
sigma_ratio=1.6,
overlap=.25,
threshold=1.75)
img_pk = np.zeros(img_mz.shape, dtype=bool)
img_pk[peaks[:, 0].astype(int), peaks[:, 1].astype(int)] = True
img_pk = morphology.label(img_pk)
# Get mask to conduct segmentation over
img_pm = self.get_primary_object_mask(
img, morphology.binary_dilation(img_pk > 0, morphology.disk(32)))
img_dt = ndi.distance_transform_edt(img_pm)
# Use propogation rather than watershed as it often captures a much more accurate boundary
img_obj = propagate.propagate(img_mz, img_pk, img_pm,
.01)[0].astype(np.uint16)
img_bnd = img_obj * segmentation.find_boundaries(
img_obj, mode='inner', background=0)
img_seg = [img_obj, img_obj, img_bnd, img_bnd]
if include_intermediate_results:
to_uint16 = lambda im: exposure.rescale_intensity(
im, out_range='uint16').astype(np.uint16)
img_seg += [
to_uint16(img_mz),
to_uint16(img_dog[0]),
to_uint16(img_dog[1]),
img_pm.astype(np.uint16),
img_pk.astype(np.uint16)
]
# Stack and add new axis to give to (z, ch, h, w)
img_seg = np.stack(img_seg)[np.newaxis]
assert img_seg.dtype == np.uint16, 'Expecting 16bit result, got type {}'.format(
img_seg.dtype)
assert img_seg.ndim == 4, 'Expecting 4D result, got shape {}'.format(
img_seg.shape)
return img_seg
def lineOrder(img, n_steps, line_thresh=50, len_thresh=60, wsz=10):
img = invert(img)
img[img <= 127] = 0
img[img > 127] = 1
i2 = img.copy() * 255
i2 = cv2.cvtColor(i2, cv2.COLOR_GRAY2RGB)
lr, lc = hough(img)
skel = skeletonize(img).astype(np.float32)
skt = skel.copy()
S = min(line_thresh, len(lr))
rlist = []
clist = []
for sel in trange(S):
lrs = []
lcs = []
resp = []
for i in range(len(lr)):
lsz = len(lr[i])
r = []
for j in range(lsz):
x = lr[i][j]
y = lc[i][j]
r.append(1 if np.sum(skt[x:x + wsz, y:y + wsz]) > 0 else 0)
r = np.array(r)
dp = np.zeros(r.shape, np.uint8)
dp[0] = r[0]
for j in range(1, r.shape[0]):
dp[j] = r[j] * (dp[j - 1] + 1)
mxi = np.argmax(dp)
lrs.append(lr[i][mxi - dp[mxi] + 1:mxi + 1])
lcs.append(lc[i][mxi - dp[mxi] + 1:mxi + 1])
resp.append(dp[mxi])
bind = np.argsort(resp)[-1]
if resp[bind] < len_thresh:
break
lre, lce = eline2(lrs[bind], lcs[bind], img, s=wsz)
lrp, lcp = eline2(lrs[bind], lcs[bind], img, s=10)
skt[lre, lce] = 0
rlist.append(lrp)
clist.append(lcp)
step = len(rlist) // n_steps
if step == 0:
step += 1
splits = list(range(0, len(rlist), step))
splits.append(len(rlist))
rans = []
cans = []
for i in range(len(splits) - 1):
rans.append(rlist[splits[i]:splits[i + 1]])
cans.append(clist[splits[i]:splits[i + 1]])
return rans, cans
def test_diameter_closing(self):
"Test for Diameter Opening (2 thresholds, all types)"
img = np.array([[97, 95, 93, 92, 91, 90, 90, 90, 91, 92, 93, 95],
[95, 93, 91, 89, 88, 88, 88, 88, 88, 89, 91, 93],
[93, 63, 63, 63, 63, 86, 86, 86, 87, 43, 43, 91],
[92, 89, 88, 86, 85, 85, 84, 85, 85, 43, 43, 89],
[91, 88, 87, 85, 84, 84, 83, 84, 84, 85, 87, 88],
[90, 88, 86, 85, 84, 83, 83, 83, 84, 85, 86, 88],
[90, 88, 86, 84, 83, 83, 82, 83, 83, 84, 86, 88],
[90, 88, 86, 85, 84, 83, 83, 83, 84, 85, 86, 88],
[91, 88, 87, 85, 84, 84, 83, 84, 84, 85, 87, 88],
[92, 89, 23, 23, 85, 85, 84, 85, 85, 3, 3, 89],
[93, 91, 23, 23, 87, 86, 86, 86, 87, 88, 3, 91],
[95, 93, 91, 89, 88, 88, 88, 88, 88, 89, 91, 93]],
dtype=np.uint8)
ex2 = np.array([[97, 95, 93, 92, 91, 90, 90, 90, 91, 92, 93, 95],
[95, 93, 91, 89, 88, 88, 88, 88, 88, 89, 91, 93],
[93, 63, 63, 63, 63, 86, 86, 86, 87, 43, 43, 91],
[92, 89, 88, 86, 85, 85, 84, 85, 85, 43, 43, 89],
[91, 88, 87, 85, 84, 84, 83, 84, 84, 85, 87, 88],
[90, 88, 86, 85, 84, 83, 83, 83, 84, 85, 86, 88],
[90, 88, 86, 84, 83, 83, 83, 83, 83, 84, 86, 88],
[90, 88, 86, 85, 84, 83, 83, 83, 84, 85, 86, 88],
[91, 88, 87, 85, 84, 84, 83, 84, 84, 85, 87, 88],
[92, 89, 23, 23, 85, 85, 84, 85, 85, 3, 3, 89],
[93, 91, 23, 23, 87, 86, 86, 86, 87, 88, 3, 91],
[95, 93, 91, 89, 88, 88, 88, 88, 88, 89, 91, 93]],
dtype=np.uint8)
ex4 = np.array([[97, 95, 93, 92, 91, 90, 90, 90, 91, 92, 93, 95],
[95, 93, 91, 89, 88, 88, 88, 88, 88, 89, 91, 93],
[93, 63, 63, 63, 63, 86, 86, 86, 87, 84, 84, 91],
[92, 89, 88, 86, 85, 85, 84, 85, 85, 84, 84, 89],
[91, 88, 87, 85, 84, 84, 83, 84, 84, 85, 87, 88],
[90, 88, 86, 85, 84, 83, 83, 83, 84, 85, 86, 88],
[90, 88, 86, 84, 83, 83, 83, 83, 83, 84, 86, 88],
[90, 88, 86, 85, 84, 83, 83, 83, 84, 85, 86, 88],
[91, 88, 87, 85, 84, 84, 83, 84, 84, 85, 87, 88],
[92, 89, 84, 84, 85, 85, 84, 85, 85, 84, 84, 89],
[93, 91, 84, 84, 87, 86, 86, 86, 87, 88, 84, 91],
[95, 93, 91, 89, 88, 88, 88, 88, 88, 89, 91, 93]],
dtype=np.uint8)
# _full_type_test makes a test with many image types.
_full_type_test(img, 2, ex2, diameter_closing, connectivity=2)
_full_type_test(img, 4, ex4, diameter_closing, connectivity=2)
P, S = max_tree(invert(img), connectivity=2)
_full_type_test(img, 4, ex4, diameter_opening,
parent=P, tree_traverser=S)
def func_images_skeletonize(images, random_state, parents, hooks):
results = []
for img in images:
# make it a binary image
img = np.squeeze(img)
thresh = threshold_otsu(img)
inv_binary = img < thresh # inverts, because normally called with `>`
inv_skeleton = morphology.skeletonize(inv_binary)
# formatting
skeleton = util.invert(inv_skeleton) * 255
skeleton = np.expand_dims(skeleton, axis=2)
skeleton = skeleton.astype(np.uint8)
results.append(skeleton)
return results
def initialize(self, opt):
self.opt = opt
self.root = opt.dataroot
### input A (label maps)
if opt.primitive != "seg_edges":
dir_A = "_" + opt.primitive
self.dir_A = os.path.join(opt.dataroot, opt.phase + dir_A)
self.A_paths = sorted(make_dataset(self.dir_A))
self.A = Image.open(self.A_paths[0])
else:
self.dir_A = os.path.join(opt.dataroot, opt.phase + "_seg")
self.A_paths = sorted(make_dataset(self.dir_A))
# the seg input will be saved as "A"
self.A = Image.open(self.A_paths[0])
self.dir_A_edges = os.path.join(opt.dataroot, opt.phase + "_edges")
if not os.path.exists(self.dir_A_edges):
os.mkdir(self.dir_A_edges)
self.A_paths_edges = sorted(make_dataset(self.dir_A_edges))
if not os.path.exists(self.dir_A_edges):
os.makedirs(self.dir_A_edges)
self.A_edges = Image.open(
self.A_paths_edges[0]) if self.A_paths_edges else None
### input B (real images)
if opt.isTrain or opt.use_encoded_image:
dir_B = '_B' if self.opt.label_nc == 0 else '_img'
self.dir_B = os.path.join(opt.dataroot, opt.phase + dir_B)
self.B_paths = sorted(make_dataset(self.dir_B))
self.B = Image.open(self.B_paths[0]).convert('RGB')
if opt.primitive == "seg_edges" and not self.A_edges:
self.A_edges = Image.fromarray(
util.invert(
feature.canny(rgb2gray(np.array(self.B)), sigma=0.5)))
self.adjust_input_size(opt)
### instance maps
if not opt.no_instance:
self.dir_inst = os.path.join(opt.dataroot, opt.phase + '_inst')
self.inst_paths = sorted(make_dataset(self.dir_inst))
### load precomputed instance-wise encoded features
if opt.load_features:
self.dir_feat = os.path.join(opt.dataroot, opt.phase + '_feat')
print('----------- loading features from %s ----------' %
self.dir_feat)
self.feat_paths = sorted(make_dataset(self.dir_feat))
self.dataset_size = len(self.A_paths)
def reverse_img(img):
"""
Take address of image to be processed as img_loc.
Computes log correction and returns both the input img and corrected img.
param:
img - address to access image to be processed
returns:
img - 2D array of input image
reverse_img - 2D array of reversed image
"""
reverse_img = np.asarray(util.invert(img), dtype='uint8')
return reverse_img[:, :, 0:3]
def fillGreen(districtName):
global year
imagenir = imread('../Imgs/LayersNIR/%s.png' % districtName, mode='RGB')
imagergb = imread('../Imgs/LayersRGB %d/%s.png' % (year, districtName),
mode='RGB')
invnir = imagenir[:, :, 0].astype('float32') / 255.0
imgred = imagergb[:, :, 0].astype('float32') / 255.0
ndvi = NDVI(imgred, util.invert(invnir))
greeniness = Greeniness(imagergb)
dmap = DMAP(imgred, invnir)
print(year, districtName, ndvi, greeniness, dmap)
return (ndvi, greeniness, dmap)
def cleanliorimage(alignedpath):
rescale = MinMaxScaler(feature_range=(0, 1))
pt = QuantileTransformer(output_distribution='normal')
img = skio.imread(alignedpath)
img = pt.fit_transform(img)
img = rescale.fit_transform(img)
img = exposure.equalize_adapthist(img, clip_limit=0.05)
imrescaled = cv2.resize(img, dsize=(1024, 1024),
interpolation=cv2.INTER_AREA)
imrescaled = util.invert(imrescaled)
imrescaled = imrescaled*(255**2)
imrescaled = imrescaled.astype('u2')
# skio.imsave(os.path.join('/'.join(alignedpath.split('/')[:-1]),"inference/")+alignedpath.split('/')[-1], imrescaled)
return(imrescaled)
def predict_classified_image(encoder_data_file_path=r"D:\encoded_data.npz",
encoder_file_path="encoder.h5",
rec_input_path=r"D:\img_test",
out_put_path=r"D:\img_out"):
'''
用于分类已经经过初步分类好的未进行人工标记的数据。由于模型已经训练好，此方法可以不用
:param encoder_data_file_path:
:param encoder_file_path:
:param rec_input_path:
:param out_put_path:
:return:
'''
labeled_data = np.load("D:/data.npz")
x_train, y_train = labeled_data['x_train'], labeled_data['y_train']
import keras
encoder = keras.models.load_model(encoder_file_path)
input_folders = os.listdir(rec_input_path)
x_train_coded = np.load(encoder_data_file_path)['arr_0']
tibetan_word_code = []
with open("words_Titan.txt") as f:
for line in f.readlines():
tibetan_word_code.append(line.strip())
for i, folder in enumerate(input_folders):
img_folder_full_path = rec_input_path + os.sep + folder
img_name = os.listdir(img_folder_full_path)[0]
rec_img_path = rec_input_path + os.sep + folder + os.sep + img_name
img = io.imread(rec_img_path, as_grey=True)
img_resize = transform.resize(util.invert(img), (60, 60),
mode='reflect')
img2 = np.reshape(img_resize, (1, 60, 60, 1))
encoded_img = encoder.predict(img2)
lbs = []
for img_t in x_train_coded:
lbs.append(cosine_similarity_2(encoded_img, img_t))
mx = np.argmax(lbs)
if lbs[mx] > 0.9:
out_classified_path = tibetan_word_code[y_train[mx]]
else:
out_classified_path = "unrecognized"
for img_file in os.listdir(img_folder_full_path):
img_full_path = os.path.join(img_folder_full_path, img_file)
img_out_put_path = os.path.join(out_put_path, out_classified_path)
if not os.path.exists(img_out_put_path):
os.mkdir(img_out_put_path)
copyfile(img_full_path, os.path.join(img_out_put_path, img_file))
print("%d of %d" % (i, len(input_folders)))
def invert_imgs(self, set=[], labels=[]):
"""Inverts each image"""
print('Inverting images...')
out_set = []
out_labels = []
for i in range(0, len(set)):
out_set.append(set[i])
out_labels.append(labels[i])
inverted_img = util.invert(set[i])
out_set.append(inverted_img)
out_labels.append(labels[i])
return np.asarray(out_set), np.asarray(out_labels)
def add_symbol_to_image(img,
folder,
choices,
padding,
minsize,
maxsize,
bpower=False,
bsmall=False,
bnom=False,
bden=False,
width=False):
choice = np.random.randint(len(choices))
symbol_img = io.imread(read_path + "/" + folder + "/" + choices[choice])
new_width = np.random.randint(minsize, maxsize + 1)
new_height = np.random.randint(minsize, maxsize + 1)
if width is not False:
new_width = width
symbol_img_res = resize(symbol_img, (new_height, new_width), cval=1) * 255
symbol_img_res = crop(symbol_img_res)
new_height, new_width = symbol_img_res.shape
shift = np.random.randint(-4 + (60 - new_height) // 2,
4 + (60 - new_height) // 2)
bounding_box = {
'xmin':
padding,
'xmax':
padding + new_width,
'ymin':
65 + shift - 15 * bpower + 10 * bsmall - 30 * bnom + 30 * bden,
'ymax':
65 + shift + new_height - 15 * bpower + 10 * bsmall - 30 * bnom +
30 * bden,
'class_text':
folder,
'class':
label_names_dict[folder]
}
if folder == "y" or folder == "beta":
bounding_box['ymin'] += 10
bounding_box['ymax'] += 10
xmin, xmax = bounding_box['xmin'], bounding_box['xmax']
ymin, ymax = bounding_box['ymin'], bounding_box['ymax']
img[ymin:ymax, xmin:xmax] += invert(symbol_img_res) + 254
padding += new_width + np.random.randint(2, 5)
return img, padding, bounding_box
def basic_circle_mask(input):
hsv = bgr2hsv(input)
msk = cv.inRange(hsv, (260 / 2, 0.1 * 255, 0.4 * 255),
(360 / 2, 255, 255)) # purples
msk += cv.inRange(hsv, (30 / 2, 0.3 * 255, 0),
(80 / 2, 255, 0.9 * 255)) # yellows
msk += cv.inRange(hsv, (0, 0, 0), (360 / 2, 255, 0.2 * 255)) # blacks
# saturated colors
window = 0.5
for s in np.arange(0.4, 1 - window, 0.1):
msk += cv.inRange(hsv, (0, s * 255, 0),
(360 / 2, 255, (s + window) * 255))
msk = ski2cv(msk > 0)
return invert(msk)
def preprocess_image(image):
"""Function that preprocess image.
Returns:
image: Preprocessed image.
"""
# invert grayscale image
image = util.invert(image)
# resize and reshape image for model
image = transform.resize(image, (28,28), anti_aliasing=True, mode="constant")
image = np.array(image)
image = image.reshape((1,28*28))
return image
def plot_image(stage, chains=None, axis_off=True, inverted=False, title=''):
if inverted:
stage = invert(stage)
fig, ax = plt.subplots(dpi=300)
ax.imshow(stage, cmap=plt.cm.gray)
if axis_off:
ax.set_axis_off()
if title:
ax.set_title(title)
if chains:
for chain in chains: # add chains to image plot
yy, xx = np.array(chain).T
plt.plot(xx, yy, lw=0.2)
plt.show()
def classifiy_chars_by_pages(path_in='/home/lyx2/img_in',
path_out=r"/home/lyx2/img_out"):
encoder_file_path = "./encoder.h5"
encoder_data_file_path = r"./encoded_data.npz"
labeled_data = np.load(r"./data.npz")
temp_folder = r'./temp__'
x_train, y_train = labeled_data['x_train'], labeled_data['y_train']
if os.path.exists(temp_folder):
rmtree(temp_folder)
os.mkdir(temp_folder)
batch_extract_char_2_file(char_out_path=temp_folder)
import keras
encoder = keras.models.load_model(encoder_file_path)
x_train_coded = np.load(encoder_data_file_path)['arr_0']
tibetan_word_code = []
with open("words_Titan.txt") as f:
for line in f.readlines():
tibetan_word_code.append(line.strip())
image_to_recog_list = os.listdir(temp_folder)
for i, image_file_name in enumerate(image_to_recog_list):
page_path = os.path.join(temp_folder, image_file_name)
t = len(os.listdir(page_path))
for j, img_name in enumerate(os.listdir(page_path)):
img_path = os.path.join(page_path, img_name)
charactor_img = io.imread(img_path, as_grey=True)
img_resize = transform.resize(util.invert(charactor_img), (60, 60),
mode='reflect')
img2 = np.reshape(img_resize, (1, 60, 60, 1))
encoded_img = encoder.predict(img2)
lbs = []
for img_t in x_train_coded:
lbs.append(cosine_similarity_2(encoded_img, img_t))
mx = np.argmax(lbs)
if lbs[mx] > 0.9:
out_classified_path = tibetan_word_code[y_train[mx]]
else:
out_classified_path = "unrecognized"
img_out_put_path = os.path.join(path_out, out_classified_path)
if not os.path.exists(img_out_put_path):
os.mkdir(img_out_put_path)
img_out_put_path = os.path.join(img_out_put_path, img_name)
copyfile(img_path, img_out_put_path)
print("Page %d, %d of %d" % (i, j, t))
rmtree(temp_folder)
def imageReady(self):
self.sendToConsole('Mouse released')
predict = self.predict
x = self.x
sess = self.sess
root_dir = self.root_dir
self.sendToConsole('Image ready')
image = QPixmap('img.png')
h = self.label_classifierInputImage.height()
w = self.label_classifierInputImage.width()
self.label_classifierInputImage.setPixmap(image.scaled(w, h))
with sess.as_default():
#load picture to be classified
temp = []
image_path = os.path.join(root_dir, 'img.png')
img = imread(image_path, flatten=True)
img = img.astype('float32')
img_28 = scipy.misc.imresize(img, (28, 28),
interp='bilinear',
mode=None)
img_inv = util.invert(img_28) #invert_color
temp.append(img_inv)
test_img = np.stack(temp)
self.sendToConsole('Classifying image...')
pred = predict.eval({x: test_img.reshape(-1, 784)})
self.sendToConsole('Classifier output: ' + str(pred) + '\n')
# save inverse 28x28image
image_inv28_save_path = os.path.join(root_dir, 'img28_inv.png')
imsave(image_inv28_save_path, img_inv)
# Load inverse 28x28image into the GUI
image = QPixmap('img28_inv.png')
h = self.label_classifierInputImage.height()
w = self.label_classifierInputImage.width()
self.label_classifierInputImage_downSampled.setPixmap(
image.scaled(w, h))
# Update the predict in the GUI
prediction = np.asscalar(pred)
self.label_classifiedDigit.setText(str(prediction))
# Update the classifier history stream
self.historyText = self.historyText + str(prediction)
self.textEdit_classificationHistory.setText(self.historyText)
self.textEdit_classificationHistory.setFocus()
self.textEdit_classificationHistory.moveCursor(QTextCursor.End)
def method1(image, gray, debug=0):
# create mask based upon red values
# Theory is that dolphins should have more red than the sea...
imgmask = createMask(image)
# probably should do something better with imgmask other than just take red channel...
# setup code such that can drop in different methods to compare effectiveness...
# estimate background then threshold it to get a mask
background = estimate_background(gray, sigma=100.)
bkgMask = invert(get_threshold(background)).astype(int)
maskArea = 1. - ((np.sum(bkgMask)) / (bkgMask.shape[0] * bkgMask.shape[1]))
if maskArea > .45:
bkgFactor = 0.0
imgmask = imgmask[:, :, 0]
else:
bkgFactor = 1.1
# combine masks
imgmask = imgmask[:, :, 0] * bkgMask
# convert to binary
imgmask = np.where(imgmask > 0, 1, 0)
# subtract background, apply mask and renormalise
backgroundSubtracted = gray - (bkgFactor * background)
if maskArea < 0.45:
backgroundSubtracted *= bkgMask
else:
backgroundSubtracted *= imgmask
backgroundSubtracted = backgroundSubtracted / np.amax(
np.abs(backgroundSubtracted))
backgroundSubtracted = img_as_ubyte(backgroundSubtracted)
if debug > 2:
labels = ["gray", "background", "backgroundSubtracted", "imgmask"]
cmaps = [plt.cm.gray for i in range(0, 4)]
figd, axs = debug_fig(gray,
background,
backgroundSubtracted,
imgmask,
labels,
cmaps,
pos=1)
return backgroundSubtracted, imgmask, axs
return backgroundSubtracted, imgmask, None
def charsegment(word, word_no, line_no):
verticalp = np.sum(word, 0)
# plt.plot(verticalp)
iszero = np.equal(verticalp, 0).view(np.int8)
# print(iszero)
absdiff = np.abs(np.diff(iszero))
# print(absdiff)
# plt.show()
try:
absdiff[-1] = 1
gaps = np.where(absdiff == 1)[0].reshape(-1, 2)
except:
absdiff[-1] = 0
gaps = np.where(absdiff == 1)[0].reshape(-1, 2)
# print(gaps)
k = 0
for i in range(len(gaps)):
crop_char = word[0:word.shape[0], gaps[i][0]:gaps[i][1] + 2]
crop_char = cv2.rotate(crop_char, cv2.ROTATE_90_CLOCKWISE)
ret, labels = cv2.connectedComponents(crop_char)
# print(labels)
props = regionprops(labels)
for prop in props:
# print(prop['label'])
cropped_shape = prop['image']
# print(cropped_shape)
cropped_shape = 255 * cropped_shape
# print(cropped_shape)
labeled_img = cv2.rotate(cropped_shape,
cv2.ROTATE_90_COUNTERCLOCKWISE)
labeled_img = np.array(labeled_img, dtype='uint8')
image = rescale(labeled_img)
image = invert(image)
images = []
image_copy = np.reshape(image, (50, 50, 1))
images.append(image_copy)
x = np.array(images)
with graph.as_default():
pre_class = model_segment.predict_classes(x)
if pre_class[0] == 1:
k = seg_connected(image, line_no, word_no, k)
else:
cv2.imwrite(
os.path.join(
img_no, 'cropchar_' + str(line_no) + '_' +
str(word_no) + '_' + str(k + 1) + '.png'), image)
k = k + 1
def stara(
smap,
circle_radius: u.deg = 100 * u.arcsec,
median_box: u.deg = 10 * u.arcsec,
threshold=6000,
limb_filter: u.percent = None,
):
"""
A method for automatically detecting sunspots in white-light data using morphological operations
Parameters
----------
smap : `sunpy.map.GenericMap`
The map to apply the algorithm to.
circle_radius : `astropy.units.Quantity`, optional
The angular size of the structuring element used in the
`skimage.morphology.white_tophat`. This is the maximum radius of
detected features.
median_box : `astropy.units.Quantity`, optional
The size of the structuring element for the median filter, features
smaller than this will be averaged out.
threshold : `int`, optional
The threshold used for detection, this will be subject to detector
degradation. The default is a reasonable value for HMI continuum images.
limb_filter : `astropy.units.Quantity`, optional
If set, ignore features close to the limb within a percentage of the
radius of the disk. A value of 10% generally filters out false
detections around the limb with HMI continuum images.
"""
data = invert(smap.data)
# Filter things that are close to limb to reduce false detections
if limb_filter is not None:
hpc_coords = sunpy.map.all_coordinates_from_map(smap)
r = np.sqrt(hpc_coords.Tx**2 + hpc_coords.Ty**2) / (
smap.rsun_obs - smap.rsun_obs * limb_filter)
data[r > 1] = np.nan
# Median filter to remove detections based on hot pixels
m_pix = int((median_box / smap.scale[0]).to_value(u.pix))
med = median(data, square(m_pix), behavior="ndimage")
# Construct the pixel structuring element
c_pix = int((circle_radius / smap.scale[0]).to_value(u.pix))
circle = disk(c_pix / 2)
finite = white_tophat(med, circle)
finite[np.isnan(finite)] = 0 # Filter out nans
return finite > threshold
def main(argv):
# Loads an image
image = io.imread(argv[0], True)
binary = image > threshold_otsu(image)
skel = skeletonize_3d(invert(binary))
# Copy edges to the images that will display the results in BGR
lines = probabilistic_hough_line(skel,
threshold=5,
line_length=20,
line_gap=50)
plot(image, skel, lines)
def renumerate(image, max=False, inver=None):
from skimage import util
if inver is not None:
image = util.invert(image)
image = rgb2gray(image)
classes = np.unique(image)
dummy = np.zeros_like(image)
for idx, value in enumerate(classes):
mask = np.where(image == value)
dummy[mask] = idx
max_class = idx
if max is True:
return dummy, max_class
if max is False:
return dummy
def get_skeleton(img):
'''
Takes in the route image on the node network
and returns a skeletonized version
Parameters:
input: grayscale image [0,1]
output: grayscale image [0,1]
'''
# img_thresh = np.where(img > .55, 1, 0)
img_thresh = img > .55
img_thresh = invert(img_thresh) + 2
skeleton = skeletonize(img_thresh) * 1.
return skeleton
def skeleton(pixels, path=0):
if type(pixels) == str:
image = io.imread(pixels, 1)
else:
image = pixels
# invertir imagen
image = invert(image)
for b, f in enumerate(image):
for a, p in enumerate(f):
image[b][a] = 0 if image[b][a] == 255 else 1
# aplicar skeletonize
skeleton = skeletonize(image)
if path != 0:
misc.imsave(path, skeleton)
return skeleton
def get_ft_image(self, idx):
if torch.is_tensor(idx):
idx = idx.tolist()
problem_name = self.df.iloc[idx]['filename']
problem_image_path = os.path.join(self.image_dir,
problem_name + '-bolded-cs.png')
image = io.imread(problem_image_path)
image = util.invert(image)
image = np.asarray(image)
image = image.astype('float32')
image = image / 255.0 # Normalize the data
transformed_image = fft2(image)
# transformed_image = transformed_image / transformed_image[0, 0]
# real_transformed_image = np.real(transformed_image) + np.imag(transformed_image)
return transformed_image
def charImageToArray(characters):
charArray = []
for character in characters:
character = util.invert(character)
character = character * 255
plt.figure()
plt.imshow(character, cmap='gray')
charArray.append(character.flatten())
# np.savetxt(dateTime+".csv", charArray, delimiter=",", fmt = "%s")
# plt.show()
return charArray
def media_axis_method(grid, grid_start, grid_goal, check_linear=True):
#check_linear specify how we do path pruning, if True, we choose colinearity
#if False, we use bresenham, usually bresenham yield a much better result
skeleton = medial_axis(invert(grid))
skel_start, skel_goal = find_start_goal(skeleton, grid_start, grid_goal)
path, _ = a_star(
invert(skeleton).astype(np.int), heuristic, tuple(skel_start),
tuple(skel_goal))
if len(path) == 0:
# print("warning, no path is found, please select another point")
return path
#insert the start and end point if necessary
if tuple(skel_start) != grid_start:
path.insert(0, grid_start)
if tuple(skel_goal) != grid_goal:
path.append(grid_goal)
#prune the path
print("path point num = {}, path={}".format(len(path), path))
path = prune_path(path, grid=grid, check_linear=check_linear)
print("pruned path point num = {}, path={}".format(len(path), path))
path = np.array(path).astype(int).tolist()
return path, skeleton, grid_start, grid_goal
def shear_image(image, slant):
""" Shear an image to correct its slant
Args:
image (matrix): The image to shear
slant (float): Slant of the image
Returns:
matrix: The sheared image
"""
# Work with inverted colors, otherwise shearing leads to black parts in image
inverted_img = util.invert(image)
# Compute shearing angle and transformation matrix
shear = math.radians(90 - float(slant))
affine_tf = transform.AffineTransform(shear=shear)
# Shear the image and crop it
inverted_img = pad_image_for_shearing(inverted_img, shear)
sheared_img = transform.warp(inverted_img, inverse_map=affine_tf)
sheared_img = crop_image_text(sheared_img)
# Invert back to white background / black text and resize to original height
sheared_img = util.invert(sheared_img)
sheared_img = resize_image(sheared_img, new_height=image.shape[0])
return sheared_img
def process(img_name, ifprint):
src = cv2.imread(img_name, cv2.IMREAD_COLOR)
if len(src.shape) != 2:
gray = cv2.cvtColor(src, cv2.COLOR_BGR2GRAY)
else:
gray = src
# robimy splot z maską, która odfiltruje poziome i pionowe linie (kratkę z kartki)
kernel = np.array([[0, 1, 0], [0, 1, 0], [0, 1, 0]])
gray2 = cv2.filter2D(gray, -1, kernel)
gray2 = invert(gray2)
gray2 = binary_dilation(gray2, disk(3))
return find_words(gray2, img_name, src, ifprint)
def predict_image_with_consine_similarity(
encoder_data_file_path=r"D:\encoded_data.npz",
encoder_file_path="encoder.h5",
rec_input_path=r"D:\Users\Riolu\Desktop\新建文件夹 (2)"):
'''
通过余弦相似度 标记 rec_input_path 中的数据
:param encoder_data_file_path: 已经使用 encoder 编码的数据
:param encoder_file_path: encoder 模型文件
:param rec_input_path: 识别的目录下的文件
:return:
# TODO 将识别的文件归类
'''
import keras
unrecognize_picture_dir = "D:\\img_out"
x_train, y_train = np.load("D:/data.npz")['x_train'], np.load(
"D:/data.npz")['y_train']
encoder = keras.models.load_model(encoder_file_path)
t = len(os.listdir(rec_input_path))
utfcode_str_list = []
x_train_coded = np.load(encoder_data_file_path)['arr_0']
for i, file_name in enumerate(os.listdir(rec_input_path)):
full_path = rec_input_path + os.path.sep + file_name
img = io.imread(full_path, as_grey=True)
img_resize = transform.resize(util.invert(img), (60, 60),
mode='reflect')
img2 = np.reshape(img_resize, (1, 60, 60, 1))
encoded_img = encoder.predict(img2)
lbs = []
for img_t in x_train_coded:
lbs.append(cosine_similarity_2(encoded_img, img_t))
mx = np.argmax(lbs)
tibetan_word_code = []
with open("words_Titan.txt") as f:
for line in f.readlines():
tibetan_word_code.append(line.strip())
if lbs[mx] > 0.9:
rec_str = tibetan_word_code[y_train[mx]]
print(rec_str, file_name)
utfcode_str_list.append(rec_str)
else:
utfcode_str_list.append('*')
io.imsave(unrecognize_picture_dir + os.path.sep + file_name, img)
# print("%d / %d" % (i,t))
print(utfcode_str_list)
def pre_image_processing(resized_image):
equal_adapt_hist_image = exposure.equalize_adapthist(resized_image)
rescale_intensity_image = exposure.rescale_intensity(equal_adapt_hist_image)
adjust_sigmoid_image = exposure.adjust_sigmoid(rescale_intensity_image)
gray_scale_image = rgb2gray(adjust_sigmoid_image)
mean_image = mean(gray_scale_image, disk(1))
mean_image = mean(mean_image, disk(1))
mean_image = mean(mean_image, disk(1))
median_image = dilation(median(mean_image, disk(1)), square(2))
otsu_image = filters.threshold_otsu(median_image)
closing_image = closing(median_image > otsu_image, square(1))
# opening_image = opening(closing_image, square(2))
opening_image = invert(closing_image)
return opening_image
def detectCirc(image):
"""
Inverts color of image and detects center of circle shape.
Asumes circle is the ONLY object in image, so noise
needs to be filtered out
"""
#staticBg = genFilter(video[0])
invFrame = image
bwFrame = gray2binary(rgb2gray(util.invert(invFrame)))[hMin:hMax,
wMin:wMax]
bwFrame = bwFrame # * staticBg
# Detects shapes in image
props = regionprops(label_image=bwFrame.astype(int))
return props, invFrame, bwFrame
def image_transformation(X, method_type='blur', **kwargs):
# https://www.kaggle.com/tomahim/image-manipulation-augmentation-with-skimage
q = kwargs['percentile'] if 'percentile' in kwargs else (0.2, 99.8)
angle = kwargs['angle'] if 'angle' in kwargs else 60
transformation_dict = {
'blur': normalize(ndimage.uniform_filter(X)),
'invert': normalize(util.invert(X)),
'rotate': rotate(X, angle=angle),
'rescale_intensity': _rescale_intensity(X, q=q),
'gamma_correction': exposure.adjust_gamma(X, gamma=0.4, gain=0.9),
'log_correction': exposure.adjust_log(X),
'sigmoid_correction': exposure.adjust_sigmoid(X),
'horizontal_flip': X[:, ::-1],
'vertical_flip': X[::-1, :],
'rgb2gray': skimage.color.rgb2gray(X)
}
return transformation_dict[method_type]
def test_invert_roundtrip():
for t, limits in dtype_range.items():
image = np.array(limits, dtype=t)
expected = invert(invert(image))
assert_array_equal(image, expected)
===========
Skeletonization reduces binary objects to 1 pixel wide representations. This
can be useful for feature extraction, and/or representing an object's topology.
``skeletonize`` works by making successive passes of the image. On each pass,
border pixels are identified and removed on the condition that they do not
break the connectivity of the corresponding object.
"""
from skimage.morphology import skeletonize
from skimage import data
import matplotlib.pyplot as plt
from skimage.util import invert
# Invert the horse image
image = invert(data.horse())
# perform skeletonization
skeleton = skeletonize(image)
# display results
fig, axes = plt.subplots(nrows=1, ncols=2, figsize=(8, 4),
sharex=True, sharey=True,
subplot_kw={'adjustable': 'box-forced'})
ax = axes.ravel()
ax[0].imshow(image, cmap=plt.cm.gray)
ax[0].axis('off')
ax[0].set_title('original', fontsize=20)
def setup(self, *args):
# we stack the horse data 5 times to get an example volume
self.image = np.stack(5 * [invert(data.horse())])
