def generate_data_NN_tagged(times,
root_dir,
NN,
NetName,
size=(128, 128),
append=True):
if append:
start_index = sum(1 for s in os.listdir(root_dir)
if s.startswith(SAMPLE_FOLDER_PREFIX))
else:
start_index = 0
for i in tqdm(range(times)):
sample_folder = os.path.join(
root_dir, SAMPLE_FOLDER_PREFIX + "{}".format(i + start_index))
try:
os.mkdir(sample_folder)
except FileExistsError:
pass
# if the cv is good enough then it will save if not it goes again
evaluation = 0
while evaluation < 0.5:
r = np.random.choice([0, 1])
shapes = ShapeGenerator(128, 128)
if r == 0:
#shapes.add_smallcorner_ellipse()
centre = shapes.add_side_ellipse()
cleanimage = shapes.image.copy()
#add some extra ellispses but not ontop of the tagged one
for i in range(np.random.choice([1, 2])):
theta = np.random.choice([90, 180])
shapes.rotation(angle=theta)
shapes.add_smallcorner_ellipse()
shapes.rotation(angle=-theta)
else:
centre = shapes.add_smallcorner_ellipse()
#shapes.add_side_ellipse()
cleanimage = shapes.image.copy()
#add some extra ellispses but not ontop of the tagged one
for i in range(np.random.choice([1, 2, 3])):
theta = np.random.choice([90, 180, 270])
shapes.rotation(angle=theta)
shapes.add_smallcorner_ellipse()
shapes.rotation(angle=-theta)
#need to trun the grey btis white for the clean segmentation
datas = cleanimage.getdata()
new_image_data = []
for item in datas:
# change all grey pixels to white
if item in list(range(50, 255)):
new_image_data.append(255)
else:
new_image_data.append(item)
# update image data
cleanimage.putdata(new_image_data)
# now all white :)
#need to randomise the corner the ellipse was placed in
theta = np.random.choice([0, 90, 180, 270])
shapes.rotation(angle=theta)
cleanimage = cleanimage.rotate(theta)
centre = rotate_around_point_highperf(centre,
theta,
origin=(shapes.height / 2,
shapes.width / 2))
#add some grey/white holes
shapes.add_holes2(
numholes=np.random.randint(40, 50),
width=np.random.randint(3, 4),
)
#add blur
shapes.add_blur(sig=1.5)
cvshapes = ChanVeseSelect(shapes.image, centre)
cvshapes.run(500, gamma=1, lmb=2, theta=0.1)
# they are meant to be reshaped inside Jaccard but python is
# ignoring that for some reason so Im doing it here :))))
u1 = np.reshape(cvshapes.u, np.size(cvshapes.u))
u2 = np.reshape(clean_seg, np.size(clean_seg))
evaluation = EM.Jaccard(u1, u2)
dsim = dst.DeepSegmentation(shapes.image, NN, cvshapes.geo, cvshapes.u)
dsim.run(1000,
lmb_reg=10,
epsilon=0.001,
gamma=10,
show_iterations=True)
# save all the images
cleanimage.save(fp=os.path.join(sample_folder,
IMAGE_TYPE_NAMES["clean"]),
format="PNG")
np.save(
file=os.path.join(sample_folder, SEGMENTATION_TYPE_NAMES["clean"]),
arr=np.array(cleanimage) / 255,
)
shapes.image.save(fp=os.path.join(sample_folder,
IMAGE_TYPE_NAMES["dirty"]),
format="PNG")
np.save(
file=os.path.join(sample_folder,
SEGMENTATION_TYPE_NAMES["chan-vese"]),
arr=cvshapes.u,
)
cvim = Image.fromarray(255 * cvshapes.u).convert("L")
cvim.save(fp=os.path.join(sample_folder,
IMAGE_TYPE_NAMES["chan-vese"]),
format="PNG")
np.save(
file=os.path.join(sample_folder, "tag.npy"),
arr=np.array(np.rint(centre).astype(int)),
)
np.save(
file=os.path.join(sample_folder, NetName + "_seg" + ".npy"),
arr=np.array(dsim.u),
)
im = Image.fromarray(255 * np.array(dsim.u)).convert("L")
im.save(fp=os.path.join(sample_folder, NetName + ".png"), format="PNG")
#save the Metrics
cv = np.reshape(cvshapes.u, np.size(cvshapes.u))
dseg = np.reshape(np.array(dsim.u), np.size(clean_seg))
cln = np.reshape(clean_seg, np.size(clean_seg))
np.save(
file=os.path.join(sample_folder, "Jaccard_" + "cv" + ".npy"),
arr=EM.Jaccard(cv, cln),
)
np.save(
file=os.path.join(sample_folder, "Jaccard_" + NetName + ".npy"),
arr=EM.Jaccard(dseg, cln),
)
np.save(
file=os.path.join(sample_folder, "Sorensen_" + "cv" + ".npy"),
arr=EM.Sorensen(cv, cln),
)
np.save(
file=os.path.join(sample_folder, "Sorensen_" + NetName + ".npy"),
arr=EM.Sorensen(dseg, cln),
)
def generate_data(times, root_dir, size=(128, 128), append=True):
if append:
start_index = sum(1 for s in os.listdir(root_dir)
if s.startswith(SAMPLE_FOLDER_PREFIX))
else:
start_index = 0
for i in tqdm(range(times)):
sample_folder = os.path.join(
root_dir, SAMPLE_FOLDER_PREFIX + "{}".format(i + start_index))
try:
os.mkdir(sample_folder)
except FileExistsError:
pass
shapes = ShapeGenerator(128, 128)
shapes.add_polygon(times=np.random.randint(10, 35))
shapes.add_ellipse(times=np.random.randint(10, 35))
shapes.add_holes(numholes=np.random.randint(5, 20),
width=np.random.randint(5, 20))
shapes.image.save(fp=os.path.join(sample_folder,
IMAGE_TYPE_NAMES["clean"]),
format="PNG")
np.save(
file=os.path.join(sample_folder, SEGMENTATION_TYPE_NAMES["clean"]),
arr=np.array(shapes.image) / 255,
)
shapes.add_noise()
shapes.image.save(fp=os.path.join(sample_folder,
IMAGE_TYPE_NAMES["dirty"]),
format="PNG")
shapes = ChanVese(shapes.image)
shapes.run(steps=500, show_iterations=False)
# save in chan-vese
np.save(
file=os.path.join(sample_folder,
SEGMENTATION_TYPE_NAMES["chan-vese"]),
arr=shapes.u,
)
def generate_data_NN(times,
root_dir,
NN,
NetName,
size=(128, 128),
append=True):
if append:
start_index = sum(1 for s in os.listdir(root_dir)
if s.startswith(SAMPLE_FOLDER_PREFIX))
else:
start_index = 0
for i in tqdm(range(times)):
sample_folder = os.path.join(
root_dir, SAMPLE_FOLDER_PREFIX + "{}".format(i + start_index))
try:
os.mkdir(sample_folder)
except FileExistsError:
pass
# if the cv is good enough then it will save if not it goes again
evaluation = 0
while evaluation < 0.5:
shapes = ShapeGenerator(128, 128)
shapes.add_ellipse(times=np.random.randint(1, 3), size=0.2 * 128)
cleanimage = shapes.image.copy()
#need to trun the grey btis white for the clean segmentation
datas = cleanimage.getdata()
new_image_data = []
for item in datas:
# change all grey pixels to white
if item in list(range(50, 255)):
new_image_data.append(255)
else:
new_image_data.append(item)
# update image data
cleanimage.putdata(new_image_data)
clean_seg = np.array(cleanimage)
# now all white :)
#add some grey/white holes
shapes.add_holes(
numholes=np.random.randint(40, 50),
width=np.random.randint(3, 4),
)
#add blur
shapes.add_blur(sig=1.5)
cvshapes = ChanVese(shapes.image)
cvshapes.run(steps=500, show_iterations=False)
# they are meant to be reshaped inside Jaccard but python is
# ignoring that for some reason so Im doing it here :))))
u1 = np.reshape(cvshapes.u, np.size(cvshapes.u))
u2 = np.reshape(clean_seg, np.size(clean_seg))
evaluation = EM.Jaccard(u1, u2)
dsim = ds.DeepSegmentation(shapes.image, NN, cvshapes.u)
dsim.run(1000, lmb_reg=10, epsilon=0.001, show_iterations=True)
# save all the images
cleanimage.save(fp=os.path.join(sample_folder,
IMAGE_TYPE_NAMES["clean"]),
format="PNG")
np.save(
file=os.path.join(sample_folder, SEGMENTATION_TYPE_NAMES["clean"]),
arr=np.array(cleanimage) / 255,
)
shapes.image.save(fp=os.path.join(sample_folder,
IMAGE_TYPE_NAMES["dirty"]),
format="PNG")
np.save(
file=os.path.join(sample_folder,
SEGMENTATION_TYPE_NAMES["chan-vese"]),
arr=cvshapes.u,
)
cvim = Image.fromarray(255 * cvshapes.u).convert("L")
cvim.save(fp=os.path.join(sample_folder,
IMAGE_TYPE_NAMES["chan-vese"]),
format="PNG")
np.save(
file=os.path.join(sample_folder, NetName + "_seg" + ".npy"),
arr=np.array(dsim.u),
)
im = Image.fromarray(255 * np.array(dsim.u)).convert("L")
im.save(fp=os.path.join(sample_folder, NetName + ".png"), format="PNG")
#save the Metrics
cv = np.reshape(cvshapes.u, np.size(cvshapes.u))
dseg = np.reshape(np.array(dsim.u), np.size(clean_seg))
cln = np.reshape(clean_seg, np.size(clean_seg))
np.save(
file=os.path.join(sample_folder, "Jaccard_" + "cv" + ".npy"),
arr=EM.Jaccard(cv, cln),
)
np.save(
file=os.path.join(sample_folder, "Jaccard_" + NetName + ".npy"),
arr=EM.Jaccard(dseg, cln),
)
np.save(
file=os.path.join(sample_folder, "Sorensen_" + "cv" + ".npy"),
arr=EM.Sorensen(cv, cln),
)
np.save(
file=os.path.join(sample_folder, "Sorensen_" + NetName + ".npy"),
arr=EM.Sorensen(dseg, cln),
)
def create(
times=1,
size=(128, 128),
cleansave="images/clean/clean_",
dirtysave="images/dirty/dirty_",
chansave="images/chan-vese/chanvese_",
datacleansaave="data/clean/clean_",
datadirtysave="data/dirty/dirty_",
datachansave="data/chan-vese/chanvese_",
):
for i in range(times):
e = datetime.datetime.now().strftime("%m_%d_%H_%M_%S_%f")
# create a clean image
shapes = ShapeGenerator(128, 128)
shapes.add_polygon(times=20)
shapes.add_ellipse(times=20)
shapes.add_holes(numholes=20)
# save in clean
shapes.image.save(fp=cleansave + e + ".png", format="PNG")
np.save(file=datacleansaave + e, arr=np.array(shapes.image) / 255)
# add noise
shapes.add_noise()
# save in dirty
shapes.image.save(fp=dirtysave + e + ".png", format="PNG")
np.save(file=datadirtysave + e, arr=np.array(shapes.image) / 255)
# apply chan-vese
shapes = ChanVese(shapes.image)
shapes.run(steps=400, show_iterations=False)
# save in chan-vese
np.save(file=datachansave + e, arr=shapes.u)
im = Image.fromarray(255 * shapes.u).convert("L")
im.save(fp=chansave + e + ".png", format="PNG")