def test_values(self):
image = cvu.load_grayscale('files/uniform.png')
stdevs = [2.5, 5, 7.5, 10]
for stdev in stdevs:
noisy = add_uniform_noise(image, stdev)
actual = np.std(noisy)
assert actual == pytest.approx(stdev, rel=0.05)
def test_dct_blocks():
image = cvu.load_grayscale('files/building_1.png')
exp_blocks, exp_means = _slow_dct_blocks(image)
act_blocks, act_means = compute_dct_blocks(image)
assert np.all(np.isclose(exp_means, act_means))
assert np.all(np.isclose(exp_blocks, act_blocks, atol=255 * 1e-5))
def test_signal_dependent_noise():
names = glob.glob('files/*.png')
for n in names:
image = cvu.load_grayscale(n)
noisy = add_signal_dependent_noise(image, 0.3, 5)
out = estimate_signal_dependent_noise(noisy, 7)
c = n.split('/')[-1].split('.')[0]
np.savetxt('out/signal_dependent/%s.txt' % c, out)
def test_patch_average():
image = cvu.load_grayscale('files/flowers_2.png')
expected = cv2.blur(image, (5, 5))
d_image = cd.to_device(image)
d_out = cd.device_array(image.shape, image.dtype)
cuda_patch_average(d_image, d_out, 2)
actual = d_out.copy_to_host()
# Except for the borders, we expect the same results
assert np.all(np.isclose(actual[2:-2, 2:-2], expected[2:-2, 2:-2]))
def test_uniform_noise_with_quantization():
names = glob.glob('files/*.png')
for n in names:
image = cvu.load_grayscale(n)
stdevs = [2.5 * i for i in range(0, 20, 3)]
estimations = []
for stdev in stdevs:
print("Estimating %s at %s" % (n, stdev))
noisy = add_uniform_noise(image, stdev, quantization=True)
estimations.append(estimate_uniform_noise(noisy))
c = n.split('/')[-1].split('.')[0]
out = np.array([stdevs, estimations])
np.savetxt('out/uniform_noise_quantization/%s.txt' % c, out)
g_means[i, j] = mean / 64
for k in range(8):
for l in range(8):
b = block[k, l]
g_blocks[i, j, k, l] = b * b
def compute_dct_blocks(image):
threadsperblock = (16, 16)
blockspergrid_x = math.ceil(image.shape[0] / threadsperblock[0])
blockspergrid_y = math.ceil(image.shape[1] / threadsperblock[1])
blockspergrid = (blockspergrid_x, blockspergrid_y)
rows, cols = image.shape
new_shape = (rows - 7, cols - 7)
d_image = cd.to_device(image)
d_means = cd.device_array(new_shape, np.float32)
d_blocks = cd.device_array((*new_shape, 8, 8), np.float32)
_kernel_compute_dct_blocks[blockspergrid,
threadsperblock](d_image, d_blocks, d_means)
return d_blocks.copy_to_host(), d_means.copy_to_host()
if __name__ == '__main__':
image = cvu.load_grayscale('files/flowers_1.png')
compute_dct_blocks(image)
def test_zero_stdev(self):
image = cvu.load_grayscale('files/woman.png')
noisy = add_uniform_noise(image, 0)
assert np.all(image == noisy)