Documentation coming soon ...
Code samples on this website come without any guarantees – use them at your own risk! I am grateful for comments or patches sent to my email address: first name.last name@cl.cam.ac.uk.
#!/usr/bin/python
import glob
import sys
import numpy
import math
import cPickle
# http://www.pybytes.com/pywavelets/ref/
import pywt
# http://www.pythonware.com/library/pil/handbook/index.htm
from PIL import Image, ImageOps
# Accelerator module for filtering.
from accel.filter_accel import wiener_filter
TILE_OVERLAP = 8
TILE_SIZE = 512
DENOISE_SIGMA = 5
def denoise_coefficient_list(coefficient_list, sigma):
ll = coefficient_list[0]
denoised_bands = [ll]
for band, subband_coefficients in enumerate(coefficient_list[1 :]):
denoised_bands.append([wiener_filter(s.astype(numpy.float), sigma)
for s in subband_coefficients])
return denoised_bands
def get_noise(greyscale_matrix):
original_shape = greyscale_matrix.shape
# The image will be transformed in TILE_SIZE * TILE_SIZE tiles with overlap
# TILE_OVERLAP on each side.
tiles_count = [(d - TILE_OVERLAP) / (TILE_SIZE - TILE_OVERLAP)
for d in original_shape]
tiled_shape = [TILE_OVERLAP + c * (TILE_SIZE - TILE_OVERLAP)
for c in tiles_count]
without_edges_shape = [c * (TILE_SIZE - TILE_OVERLAP) - TILE_OVERLAP
for c in tiles_count]
# The greyscale image is represented as a matrix of float values.
greyscale_matrix = greyscale_matrix.astype(float)
result_matrix = numpy.zeros(tiled_shape, dtype = numpy.float)
# Work out how many levels of wavelet decomposition we will do.
dyad_length = math.ceil(math.log(TILE_SIZE, 2))
ll_levels = 5
wavelet_levels = dyad_length - ll_levels
ll_size = 2 ** ll_levels
# Make a window for the tile edges.
tile_window = numpy.zeros((TILE_SIZE, TILE_SIZE), dtype=numpy.float)
tile_window[TILE_OVERLAP / 2 :
-(TILE_OVERLAP / 2),
TILE_OVERLAP / 2 :
-(TILE_OVERLAP / 2)] = 1.0
# Transform and filter each non-overlapping TILE_SIZE * TILE_SIZE square of
# the image separately.
for ty in range(0, tiles_count[1]):
for tx in range(0, tiles_count[0]):
print (tx, ty)
transform_input = greyscale_matrix[
tx * (TILE_SIZE - TILE_OVERLAP) :
tx * (TILE_SIZE - TILE_OVERLAP) + TILE_SIZE,
ty * (TILE_SIZE - TILE_OVERLAP) :
ty * (TILE_SIZE - TILE_OVERLAP) + TILE_SIZE]
coefficient_list = pywt.wavedec2(transform_input,
'db8',
level = int(wavelet_levels),
mode = 'per')
coefficient_list = denoise_coefficient_list(coefficient_list,
DENOISE_SIGMA)
denoised_tile = pywt.waverec2(coefficient_list,
'db8',
mode='per')
denoised_tile[denoised_tile > 255.0] = 255.0
denoised_tile[denoised_tile < 0.0] = 0.0
result_matrix[tx * (TILE_SIZE - TILE_OVERLAP) :
tx * (TILE_SIZE - TILE_OVERLAP) + TILE_SIZE,
ty * (TILE_SIZE - TILE_OVERLAP) :
ty * (TILE_SIZE - TILE_OVERLAP) + TILE_SIZE] += \
(denoised_tile * tile_window)
# Remove the edges.
result_matrix = result_matrix[TILE_OVERLAP : -TILE_OVERLAP,
TILE_OVERLAP : -TILE_OVERLAP]
Image.fromstring('L',
(result_matrix.shape[1], result_matrix.shape[0]),
(result_matrix / 2.0)
.astype(numpy.uint8)
.tostring()).save('denoisedtest.png', 'PNG')
# Subtract the denoised image from the original to get an estimate of the
# noise.
result_matrix = greyscale_matrix[
TILE_OVERLAP : tiled_shape[0] - TILE_OVERLAP,
TILE_OVERLAP : tiled_shape[1] - TILE_OVERLAP] \
- result_matrix
return result_matrix
def get_noise_from_file(file_name):
original = Image.open(file_name)
greyscale = ImageOps.grayscale(original)
greyscale_vector = numpy.fromstring(greyscale.tostring(), dtype=numpy.uint8)
greyscale_matrix = numpy.reshape(greyscale_vector,
(original.size[1], original.size[0]))
noise_matrix = get_noise(greyscale_matrix)
return noise_matrix
# Command line utility for creating the characteristic.
if __name__ == '__main__':
if len(sys.argv) != 2:
print "Usage:\n\t%s path_with_png_files" % (sys.argv[0],)
sys.exit(0)
# Get a list of images to process.
file_list = glob.glob(sys.argv[1] + '/*.png')
print "Processing %d images" % (len(file_list),)
# Denoise each image, and add the noise to the average_buffer.
average_buffer = None
for i, f in enumerate(file_list):
print "Processing %03d %s" % (i, f,)
noise_matrix = get_noise_from_file(f)
if average_buffer == None:
average_buffer = numpy.zeros_like(noise_matrix)
average_buffer += noise_matrix
# Dump the average buffer to a file.
numpy.savetxt('noise_data.dat', average_buffer)
Image.fromstring('L',
(average_buffer.shape[1], average_buffer.shape[0]),
((255.0 + (average_buffer / (i + 1))) / 2.0)
.astype(numpy.uint8)
.tostring()).save('noise%03d.png' % (i,), 'PNG')
#!/usr/bin/python
#
# test_characteristic.py
from PIL import Image, ImageOps
from make_characteristic import get_noise_from_file
import numpy
import cPickle
import glob
import sys
TILE_OVERLAP = 8
if len(sys.argv) != 3:
print "Usage:\n\t%s noise_data.dat path_with_png_files" % (sys.argv[0],)
sys.exit(0)
noise_file_name = sys.argv[1]
image_path_name = sys.argv[2]
# Load the camera noise.
camera_noise = numpy.loadtxt(noise_file_name, dtype=numpy.float)
camera_noise_average = numpy.average(camera_noise)
camera_noise -= camera_noise_average
camera_noise_norm = numpy.sqrt(numpy.sum(camera_noise * camera_noise))
file_list = glob.glob(image_path_name + '/*.png')
print "Processing %d images" % (len(file_list),)
for f in file_list:
# Get this image's noise.
image_noise = get_noise_from_file(f)
image_noise_average = numpy.average(image_noise)
image_noise -= image_noise_average
image_noise_norm = numpy.sqrt(numpy.sum(image_noise * image_noise))
# Calculate the correlation between the two signals.
print "Dot product %s is: %s" % (f,
numpy.sum(camera_noise * image_noise) /
(camera_noise_norm * image_noise_norm))
# accel/setup.py
#
# Compile the accelerator module with python setup.py build_ext
from distutils.core import setup
from distutils.extension import Extension
from Cython.Distutils import build_ext
import numpy
ext_modules = [Extension("filter_accel",
["filter_accel.pyx"],
extra_compile_args = ['-I%s' %
(numpy.get_include(),)])]
setup(name = 'Filtering accelerator',
cmdclass = {'build_ext' : build_ext},
ext_modules = ext_modules)
# accel/filter_accel.pyx
cimport numpy
import numpy
DTYPE = numpy.float
ctypedef numpy.float_t DTYPE_t
cdef inline int int_max(int a, int b): return a if a >= b else b
cdef inline int int_min(int a, int b): return a if a <= b else b
def wiener_filter(numpy.ndarray[DTYPE_t, ndim=2]
subband_coefficients not None,
int sigma):
cdef int \
y, x, w, l, k, xmin, xmax, ymin, ymax
assert subband_coefficients.dtype == DTYPE
cdef numpy.ndarray[DTYPE_t, ndim=2] \
result_coefficients = numpy.zeros_like(subband_coefficients)
(xmin, xmax, ymin, ymax) = (0,
0,
subband_coefficients.shape[0] - 1,
subband_coefficients.shape[1] - 1)
sigma_squared = float(sigma * sigma)
for y in range(0, subband_coefficients.shape[1]):
for x in range(0, subband_coefficients.shape[0]):
variances = []
for w in [3, 5, 7, 9]:
accumulator = 0.0
# The sub-band is padded by repetition.
for l in range(-(w + 1) / 2, (w + 1) / 2 + 1):
for k in range(-(w + 1) / 2, (w + 1) / 2 + 1):
ki = int_min(int_max(x + k, xmin), xmax)
li = int_min(int_max(y + l, ymin), ymax)
accumulator += subband_coefficients[ki, li] * \
subband_coefficients[ki, li] \
- sigma_squared
accumulator /= float(w * w)
variances.append(max(0.0, accumulator))
minimum_local_variance = min(variances)
result_coefficients[x, y] = subband_coefficients[x, y] * \
(minimum_local_variance / \
(minimum_local_variance + sigma_squared))
return result_coefficients