def augmentation(self, imgs):
s0, s1, s2, gt = imgs
# reconstruct cosine curve
i_0, i_45, i_90, i_135 = calc_fourPolar_from_stokes(s0, s1, s2)
i_list = np.array([i_0, i_45, i_90, i_135])
# add gaussian noise
row, col, ch = s0.shape
for inum in range(i_list.shape[0]):
gauss = np.random.normal(0, self.sigma, (row, col, ch))
gauss = gauss.reshape(row, col, ch)
i_list[inum] += gauss
# calc new stokes images
i_0 = i_list[0]
i_45 = i_list[1]
i_90 = i_list[2]
i_135 = i_list[3]
s0, s1, s2 = calc_s0s1s2_from_fourPolar(i_0, i_45, i_90, i_135)
s0 = np.clip(s0, 0, 1)
s1 = np.clip(s1, -1, 1)
s2 = np.clip(s2, -1, 1)
s0, s1, s2 = de_normalize_s0s1s2(s0, s1, s2, self.max)
return (s0, s1, s2, gt)
def operation(self, s0, s1, s2):
i_0, i_45, i_90, i_135 = pol.calc_fourPolar_from_stokes(s0, s1, s2)
i_0 = self.averageRGB(i_0) * self.range_max
i_45 = self.averageRGB(i_45) * self.range_max
i_90 = self.averageRGB(i_90) * self.range_max
i_135 = self.averageRGB(i_135) * self.range_max
return np.array([i_0, i_45, i_90, i_135])
def operation(self, s0, s1, s2):
i_0, i_45, i_90, i_135 = pol.calc_fourPolar_from_stokes(s0, s1, s2)
i_0 *= self.range_max
i_45 *= self.range_max
i_90 *= self.range_max
i_135 *= self.range_max
stokes_array = np.array([
i_0[2, :, :], i_0[1, :, :], i_0[0, :, :], i_45[2, :, :],
i_45[1, :, :], i_45[0, :, :], i_90[2, :, :], i_90[1, :, :],
i_90[0, :, :], i_135[2, :, :], i_135[1, :, :], i_135[0, :, :]
])
return stokes_array
def augmentation(self, imgs):
s0, s1, s2, gt, s0spe, s1spe, s2spe, s0dif, s1dif, s2dif = imgs
"""mn = np.mean(s0) * self.max
newmn = np.random.normal(mn, mn/4)
gain = newmn / mn
if gain < 0:
raise ValueError("gain should be larger than 0!")"""
# random number [a, b)
a, b = 0.5, 1.5
gain = (b - a) * np.random.rand() + a
# reconstruct cosine curve
i_0, i_45, i_90, i_135 = calc_fourPolar_from_stokes(s0, s1, s2)
i_0spe, i_45spe, i_90spe, i_135spe = calc_fourPolar_from_stokes(
s0spe, s1spe, s2spe)
i_0dif, i_45dif, i_90dif, i_135dif = calc_fourPolar_from_stokes(
s0dif, s1dif, s2dif)
# multiply gain
i_0 *= gain
i_45 *= gain
i_90 *= gain
i_135 *= gain
i_0spe *= gain
i_45spe *= gain
i_90spe *= gain
i_135spe *= gain
i_0dif *= gain
i_45dif *= gain
i_90dif *= gain
i_135dif *= gain
# return back to s0s1s2
s0 = i_0 + i_90
s1 = i_0 - i_90
s2 = i_45 - i_135
s0spe = i_0spe + i_90spe
s1spe = i_0spe - i_90spe
s2spe = i_45spe - i_135spe
s0dif = i_0dif + i_90dif
s1dif = i_0dif - i_90dif
s2dif = i_45dif - i_135dif
# clipping
s0 = np.clip(s0, 0, 1)
s1 = np.clip(s1, -1, 1)
s2 = np.clip(s2, -1, 1)
s0spe = np.clip(s0spe, 0, 1)
s1spe = np.clip(s1spe, -1, 1)
s2spe = np.clip(s2spe, -1, 1)
s0dif = np.clip(s0dif, 0, 1)
s1dif = np.clip(s1dif, -1, 1)
s2dif = np.clip(s2dif, -1, 1)
# denormalize
s0, s1, s2 = de_normalize_s0s1s2(s0, s1, s2, self.max)
s0spe, s1spe, s2spe = de_normalize_s0s1s2(s0spe, s1spe, s2spe, self.max)
s0dif, s1dif, s2dif = de_normalize_s0s1s2(s0dif, s1dif, s2dif, self.max)
return (s0, s1, s2,
gt,
s0spe, s1spe, s2spe,
s0dif, s1dif, s2dif)
for gain in gains:
data = mydataAugmentation_difSpe.DataAugmentation_difSpe_Base(
csvpaths, 16, "gain_{}".format(str(gain)))
for path in data.path_list:
data.make_savedir(path)
savepath = pathlib.Path(str(path) + "_" + data.operationName)
csv = np.loadtxt(path.joinpath("path.csv"), delimiter=",", dtype=str)
for i in range(csv.shape[0]):
s0, s1, s2, gt, s0spe, s1spe, s2spe, s0dif, s1dif, s2dif = data.load_image(
path, csv[i, :])
i0, i45, i90, i135 = mpl.calc_fourPolar_from_stokes(s0, s1, s2)
i0_spe, i45_spe, i90_spe, i135_spe = mpl.calc_fourPolar_from_stokes(
s0spe, s1spe, s2spe)
i0_dif, i45_dif, i90_dif, i135_dif = mpl.calc_fourPolar_from_stokes(
s0dif, s1dif, s2dif)
# multiply gain
i0 *= gain
i45 *= gain
i90 *= gain
i135 *= gain
i0_spe *= gain
i45_spe *= gain
i90_spe *= gain
i135_spe *= gain