def lpf(image, sigma, mode=2):
(Mx, My) = image.shape
if mode == 1:
kernel = matlab_style_gauss2D(image.shape, sigma)
kernel /= numpy.max(kernel)
if Mx == 1 or My == 1:
fft = numpy.fft.fft(image)
fft = numpy.fft.fftshift(fft)
fft *= kernel
result = numpy.real(numpy.fft.ifft(numpy.fft.fftshift(fft)))
else:
fft = numpy.fft.fftshift(numpy.fft.fft2(image))
fft *= kernel
result = numpy.real(numpy.fft.ifft2(numpy.fft.fftshift(fft)))
elif mode == 2:
new_dim = 2 * array(image.shape)
kernel = matlab_style_gauss2D((new_dim[0], new_dim[1]), sigma * 2)
kernel /= numpy.max(kernel)
kernel = kernel[Mx:, My:]
image = image.astype(numpy.double)
if Mx == 1 or My == 1:
dct = fftpack.dct(image, type=1)
dct *= kernel
result = numpy.real(fftpack.idct(dct, type=1))
else:
dct = fftpack.dct(fftpack.dct(image.T, type=2, norm='ortho').T, type=2, norm='ortho')
dct *= kernel
result = numpy.real(fftpack.idct(fftpack.idct(dct.T, type=2, norm='ortho').T, type=2, norm='ortho'))
return result
def process_cube(img, weight, quality):
# TODO: check to make sure that size of img, and q_tables are consistent
img = img.copy()
# print('process_cube input: {}'.format(img))
this_quality = np.round(np.max(weight)*quality)
if this_quality < 0:
this_quality = 0
if this_quality > quality - 1:
this_quality = quality - 1
for i in range(img.shape[3]):
img[:, :, :, i] = cv2.cvtColor(img[:, :, :, i], cv2.COLOR_BGR2LAB)
img = np.float32(img)
# print('process_cube pre DCT: {}'.format(img))
# img_dct = dct(dct(dct(img, axis=0)/4, axis=1)/4, axis=3)/4
img_dct = dct(dct(img, axis=0)/4, axis=1)/4
Q_luma = luminance_tables[:, :, :, this_quality].astype(np.float32)
Q_chroma = chrominance_tables[:, :, :, this_quality].astype(np.float32)
# Q_luma[:, :, :] = .01
# Q_chroma[:, :, :] = .01
# print('Q_luma: {}'.format(Q_luma))
# print('Q_chroma: {}'.format(Q_chroma))
# print('dct, pre rounding: {}'.format(img_dct))
img_dct[:, :, 0, :] /= Q_luma
img_dct[:, :, 1, :] /= Q_chroma
img_dct[:, :, 2, :] /= Q_chroma
img_dct = np.round(img_dct)
img_dct[:, :, 0, :] *= Q_luma
img_dct[:, :, 1, :] *= Q_chroma
img_dct[:, :, 2, :] *= Q_chroma
# print('dct, post rounding: {}'.format(img_dct))
# img_processed = idct(idct(idct(img_dct, axis=0)/4, axis=1)/4, axis=3)/4
img_processed = idct(idct(img_dct, axis=0)/4, axis=1)/4
# print('process_cube post DCT: {}'.format(img_processed))
img_processed = np.clip(img_processed, 0, 255)
img_processed = np.uint8(img_processed)
for i in range(img.shape[3]):
img_processed[:,:,:,i] = cv2.cvtColor(img_processed[:,:,:,i], cv2.COLOR_LAB2BGR)
# print('process_cube output: {}'.format(img))
# print('pre dct / post_dct: {}'.format(pre_dct / post_dct))
return img_processed
def idct2(arr):
"""
@params { np.ndarray } arr
@return { np.ndarray }
"""
array = np.float64(arr)
result = idct(idct(array, axis=0), axis=1)
return result
def LSDecompFW(wav, width= 16384, max_nnz_rate=8000.0/262144.0, sparsify = 0.01, taps = 10,
level = 3, wl_weight = 1, verbose = False,fc=120):
MaxiterA = 60
length = len(wav)
n = sft.next_fast_len(length)
signal = np.zeros((n))
signal[0:length] = wav[0:length]
h0,h1 = daubcqf(taps,'min')
L = level
#print(n)
original_signal = lambda s: sft.idct(s[0:n]) + (1.0)*(wl_weight)*idwt(s[n+1:],h0,h1,L)[0]
LSseparate = lambda x: np.concatenate([sft.dct(x),(1.0)*(wl_weight)*dwt(x,h0,h1,L)[0]],axis=0)
#measurment
y = signal
#FISTA
###############################
cnnz = float("Inf")
c = signal
temp = LSseparate(y)
temp2 = original_signal(temp)
print('aaa'+str(temp2.shape))
maxabsThetaTy = max(abs(temp))
while cnnz > max_nnz_rate * n:
#FISTA
tau = sparsify * maxabsThetaTy
tolA = 1.0e-7
fh = (original_signal,LSseparate)
c = relax.fista(A=fh, b=y,x=LSseparate(c),tol=tolA,l=tau,maxiter=MaxiterA )[0]
cnnz = np.size(np.nonzero(original_signal(c)))
print('nnz = '+ str(cnnz)+ ' / ' + str(n) +' at tau = '+str(tau))
sparsify = sparsify * 2
if sparsify == 0.166:
sparsify = 0.1
signal_dct = sft.idct(c[0:n])
signal_dct = signal_dct[0:length]
signal_wl = (1.0) * float(wl_weight) * idwt(c[n+1:],h0,h1,level)[0]
signal_wl = signal_wl[0:length]
return signal_dct,signal_wl
def laplacian_pca_TV(res, x, f0, lam, gam, iter = 10):
'''
TV version of Laplacian embedding
:param res: resolution of the grid
:param x: numpy array of data in rows
:param f0: initial embedding matrix
:param lam: sparsity parameter
:param gam: fidelity parameter
:param iter: number of iterations to carry out
:return: returns embedding matrix
'''
# f0 is an initial projection
n = res ** 2
num_data = x.shape[0]
D = sparse_discrete_diff(res)
M = 1/(lam*laplacian_eigenvalues(res).reshape(n)+gam)
f = f0
y = x .dot(f)
z = shrink(y .dot(D.T), lam)
for i in range(iter):
# Update z
z_old = z
z = shrink(y .dot (D.T), lam)
# Update f
f_old = f
u, s, v = la.svd(x.T .dot (y), full_matrices=False)
f = u .dot(v)
# Update y
y_old = y
q = lam * z .dot (D) + gam * x .dot(f)
# print('norm of y before is %f' % np.sum(q ** 2))
y = fftpack.dct(q.reshape((num_data, res, res)), norm='ortho') # Images unraveled as rows
y = fftpack.dct(np.swapaxes(y,1,2), norm='ortho') # Swap and apply dct on the other side
# print('norm of y after is %f' % np.sum(y ** 2))
y = np.apply_along_axis(lambda v: M * v, 1, y.reshape((num_data, n)))
y = fftpack.idct(y.reshape((num_data, res, res)), norm='ortho')
y = fftpack.idct(np.swapaxes(y,1,2), norm='ortho')
y = y.reshape((num_data, n))
zres = np.sqrt(np.sum((z - z_old) ** 2))
znorm = np.sqrt(np.sum((z)**2))
yres = np.sqrt(np.sum((y - y_old) ** 2))
ynorm = np.sqrt(np.sum((y)**2))
fres = np.sqrt(np.sum((f - f_old) ** 2))
value = np.sum(abs(z)) + 0.5*lam*np.sum((z-y .dot(D.T))**2) + 0.5*gam*np.sum((y- x .dot(f) )**2)
print('Iter %d Val %f Z norm %f Z res %f Ynorm %f Y res %f F res %f' % (i, value, znorm, zres, ynorm, yres, fres))
return f
def idct(self, arr,coef_size=0):
if(coef_size==0):
idcta2 = fftpack.idct(arr,norm='ortho')
return idcta2
arrlen = len(arr)
if coef_size > arrlen:
new_arr = self.interpolation_with_zeros(arr,coef_size)
else:
new_arr = arr[0:int(coef_size)]
idcta2 = fftpack.idct(new_arr,norm='ortho')
return idcta2
def razafindradina_embed(grayscale_container_path, grayscale_watermark_path, watermarked_image_path, alpha):
"""
Razafindradina embedding method implementation.
Outputs the resulting watermarked image
23-July-2015
"""
grayscale_container_2darray = numpy.asarray(Image.open(grayscale_container_path).convert("L"))
grayscale_watermark_2darray = numpy.asarray(Image.open(grayscale_watermark_path).convert("L"))
assert (
(grayscale_container_2darray.shape[0] == grayscale_container_2darray.shape[1])
and (grayscale_container_2darray.shape[0] == grayscale_watermark_2darray.shape[0])
and (grayscale_container_2darray.shape[1] == grayscale_watermark_2darray.shape[1])
), "GrayscaleContainer and GrayscaleWatermark sizes do not match or not square"
# Perform DCT on GrayscaleContainer
# print grayscale_container_2darray
gcdct = dct(dct(grayscale_container_2darray, axis=0, norm="ortho"), axis=1, norm="ortho")
# print grayscale_container_2darray
# Perform SchurDecomposition on GrayscaleWatermark
gwsdt, gwsdu = schur_decomposition(grayscale_watermark_2darray)
# alpha-blend GrayscaleWatermark TriangularMatrix into GrayscaleContainer DCT coeffs with alpha
gcdct += gwsdt * alpha
# Perform IDCT on GrayscaleContainer DCT coeffs to get WatermarkedImage
watermarked_image_2darray = idct(idct(gcdct, axis=0, norm="ortho"), axis=1, norm="ortho")
watermarked_image_2darray[watermarked_image_2darray > 255] = 255
watermarked_image_2darray[watermarked_image_2darray < 0] = 0
watermarked_image = Image.fromarray(numpy.uint8(watermarked_image_2darray))
# watermarked_image.show()
# Write image to file
watermarked_image.save(watermarked_image_path)
return
def inverse_transform(data):
result = []
for i in range(len(data[0])):
result.append([])
for i in range(len(data)):
partial_result = idct(data[i])
for j in range(len(partial_result)):
result[j].append(partial_result[j])
final_result = []
for i in range(len(result)):
partial_result = idct(result[i])
final_result.append(partial_result)
return final_result
def idctii(x, axes=None):
"""
Compute a multi-dimensional inverse DCT-II over specified array axes.
This function is implemented by calling the one-dimensional inverse
DCT-II :func:`scipy.fftpack.idct` with normalization mode 'ortho'
for each of the specified axes.
Parameters
----------
a : array_like
Input array
axes : sequence of ints, optional (default None)
Axes over which to compute the inverse DCT-II.
Returns
-------
y : ndarray
Inverse DCT-II of input array
"""
if axes is None:
axes = list(range(x.ndim))
for ax in axes[::-1]:
x = fftpack.idct(x, type=2, axis=ax, norm='ortho')
return x
def convolveGaussianDCT(x, sigma, pad_sigma=4, mode="same", cache={}):
"""
1D convolution of x with Gaussian of width sigma pixels
If pad_sigma>0, pads ends with zeros by int(pad_sigma*sigma) pixels
Otherwise does unpadded fast cosine transform, hence reflection from the ends
"""
fill = int(pad_sigma * sigma)
actual_size = x.size + fill * 2
if fill > 0:
s = nearestFFTnumber(actual_size)
fill2 = s - x.size - fill
padded_x = np.pad(x, (fill, fill2), mode="constant")
else:
padded_x = x
s = padded_x.size
hnorm = sigma / float(s)
gauss = cache.get((s, hnorm))
if gauss is None:
gauss = np.exp(-(np.arange(0, s) * (np.pi * hnorm)) ** 2 / 2.0)
cache[(s, hnorm)] = gauss
res = fftpack.idct(fftpack.dct(padded_x, overwrite_x=fill > 0) * gauss, overwrite_x=fill > 0) / (2 * s)
if fill == 0:
return res
if mode == "same":
return res[fill:-fill2]
elif mode == "valid":
return res[fill * 2 : -fill2 - fill]
else:
raise ValueError("mode not supported for convolveGaussianDCT")
def partialChebyshev(field, order=1, halfLength=None):
if halfLength is None:
halfLength = 1.0;
if(len(shape(field)) == 2):
length,breadth = shape(field);
else:
length = len(field);
breadth = 1;
print length;
trans = dct(field,type=1,axis=0);
coeff = arange(length,dtype=float);
temp = zeros(shape(field), dtype=float);
trans = 2.0*trans*coeff;
temp[:,length-2] = trans[:,length-1];
for i in arange(length-2,0,-1):
temp[:,i-1] = temp[:,i+1] + trans[:,i]
return idct(temp, type=1,axis=0)/(2.0*(length-1)*halfLength);
def spectrogramToSnippet(spectrogram):
snippet = np.zeros(shape = (spectrogram.shape[0]* spectrogram.shape[1]))
num_windows = spectrogram.shape[0]
for i in range(num_windows):
idx = [i*window_len, (i+1)*window_len]
snippet[idx[0], idx[1]] = idct(windows[i, :])
return snippet
def icosine_transform(data):
w, h = data.shape
image = data.copy()
for i in range(w):
image[i][:] = idct(data[i][:], norm='ortho')
return image
def write_to_image(path, text):
x1, x2, x3, y1, y2, y3 = [4, 5, 3, 3, 5, 4]
index = 0
D = 5
img = Image.open(path)
img.getdata()
bitext = text_to_binary(text)
bitext = bitext + '0000000000000000'
r, g, b = [np.array(x) for x in img.split()]
lx, ly = r.shape()
for x in xrange(0, lx - 2, 8):
for y in xrange(0, ly - 2, 8):
if index == len(bitext) - 1:
break
metric = r[x:x + 8, y:y + 8].astype('float')
metric = dct(metric, norm='ortho')
if bitext[index] == 1:
metric[x1, y1] = max(metric[x1, y1], metric[x3, y3] + D + 1)
metric[x2, y2] = max(metric[x2, y2], metric[x3, y3] + D + 1)
else:
metric[x1, y1] = min(metric[x1, y1], metric[x3, y3] - D - 1)
metric[x2, y2] = min(metric[x2, y2], metric[x3, y3] - D - 1)
index = index + 1
metric = idct(metric, norm='ortho')
r[x:x + 8, y:y + 8] = metric.astype('uint8')
im = Image.merge("RGB", [Image.fromarray(x) for x in [r, g, b]])
im.save('%s_writed' % path)
def decompress(self, qmat, block_size,
point_count, interval, min_ts):
"""Decompresses the bitarray and returns a Dataset object."""
coeffs = zeros(block_size)
values = zeros(point_count)
ind = 0
i = 0
block_num = 0
while ind < self.bits.length():
if qmat[i] == 52:
coeffs[i] = 0.
elif ind > len(self.bits) - (64 - qmat[i]):
# File is over
break
else:
v = self.bits[ind:ind + 64 - qmat[i]]
v.extend(qmat[i] * (False,))
coeffs[i] = struct.unpack(">d", v.tobytes())[0]
ind += 64 - qmat[i]
i += 1
if i >= block_size:
values[block_num * block_size:block_num * block_size + block_size] = idct(coeffs, norm='ortho')
i = 0
block_num += 1
# We pad out to a full byte at the end of the block
if ind % 8 != 0:
ind += 8 - (ind % 8)
if i > 1:
raise SystemExit
return Dataset(point_count, interval, min_ts, values)
def outlier_removal_smoothing_dct(tsData, n):
"""
>>> data = pd.read_csv('input.csv')
>>> outlier_removal_smoothing_dct(data, 15)
{'buy': 28833387.903209731, 'sale': 40508532.108399086, 'method': 'outlier_removal_smoothing_dct'}
>>> outlier_removal_smoothing_dct(data, 20)
{'buy': 30315166.377325296, 'sale': 42088164.543017924, 'method': 'outlier_removal_smoothing_dct'}
"""
tsData['ingestdatetime'] = tsData.apply(lambda row: datetime.strptime(str(row['ingestdate']), "%Y%m%d"), axis=1)
tsData = tsData.sort(['ingestdate'], ascending=[1])
N = len(tsData)
t = range(len(tsData))
x = [float(elem) for elem in tsData['qtyavail']]
y = dct(x, norm='ortho')
window = np.zeros(N)
window[:n] = 1
yr = idct(y*window, norm='ortho')
diffs = np.diff(yr)
result_sale = abs(sum(filter(lambda x: x < 0, diffs)))
result_buy = abs(sum(filter(lambda x: x > 0, diffs)))
result = {}
result['sale'] = result_sale
result['buy'] = result_buy
result['method'] = inspect.stack()[0][3]
return result
def on_adjust(self, event=None):
"""
Executed when slider changes value
"""
width, height = self.lena.shape
quant = self.sld.GetValue()
# reconstructed
rec = np.zeros(self.lena.shape, dtype=np.int64)
for y in xrange(0,height,8):
for x in xrange(0,width,8):
d = self.lena[y:y+8,x:x+8].astype(np.float)
D = dct(dct(d.T, norm='ortho').T, norm='ortho').reshape(64)
Q = np.zeros(64, dtype=np.float)
Q[self.unzig[:quant]] = D[self.unzig[:quant]]
Q = Q.reshape([8,8])
q = np.round(idct(idct(Q.T, norm='ortho').T, norm='ortho'))
rec[y:y+8,x:x+8] = q.astype(np.int64)
diff = np.abs(self.lena-rec)
im = self.axes.flat[0].imshow(self.lena, cmap='gray')
self.axes.flat[0].set_title('Original Image')
im = self.axes.flat[1].imshow(rec, cmap='gray')
self.axes.flat[1].set_title('Reconstructed Image')
self.hinton(ax=self.axes.flat[2])
self.axes.flat[2].set_title('DCT Coefficient Mask')
im = self.axes.flat[3].imshow(diff, cmap='hot', vmin=0, vmax=255)
self.axes.flat[3].set_title('Error Image')
for ax in self.axes.flat:
ax.axis('off')
self.fig.subplots_adjust(right=0.8)
cbar_ax = self.fig.add_axes([0.85, 0.15, 0.05, 0.7])
self.fig.colorbar(im, cax=cbar_ax)
p = self.psnr(self.lena, rec)
s = self.compute_ssim(self.lena, rec)
self.statusbar.SetStatusText('PSNR :%.4f SSIM: %.4f' % (p, s))
self.canvas.draw()
def _fit_idct(self,func,N):
""" Return the chebyshev coefficients using the idct """
pts = -cheb.chebpts1(N)
y = func(pts)
coeffs = idct(y,type=3)/N
coeffs[0] /= 2.
coeffs[-1] /= 2.
return coeffs
def idct(c):
"""
Apply inverce DCT to all elements in b
:param c:
:return:
"""
return fftpack.idct(c, norm='ortho')
def foreach_idct(self):
"""
Выполняет обратное дискретное преобразование косинусов
:return:
"""
for i in range(self.nrows):
for j in range(self.ncols):
self.image_blocks[i, j, ...] = idct(np.transpose(self.image_blocks_magnitude[i, j, ...]),
**self.dct_params)
def poisson_neumann(gx, gy):
height, width = gx.shape
f = weighted_finite_differences(gx, gy, np.ones_like(gx), np.ones_like(gy))
# Compute cosine transform
fcos = dct(dct(f, norm='ortho', axis=0), norm='ortho', axis=1)
# Compute the solution in the fourier domain
x, y = np.meshgrid(np.arange(width), np.arange(height))
denom = (2.0 * np.cos(np.pi * x / width) - 2.0) + (2.0 * np.cos(np.pi * y / height) - 2.0)
# First element is fixed to 0.0
fcos /= denom
fcos[0, 0] = 0
# Compute inverse discrete cosine transform to find solution
# in spatial domain
return idct(idct(fcos, norm='ortho', axis=0), norm='ortho', axis=1)
def trans(a):
global qa
b = dct(a,type=2,n=None,axis=-1,norm='ortho',overwrite_x=False)
c = idct(b, type=2, n=None, axis=-1, norm='ortho', overwrite_x=False)
F = np.zeros((8,8))
for i in range(8):
for j in range(8):
F[i][j] = b[i][j]/qa[i][j]
return F
def background_reconstruction_dct(I):
"""
background_reconstruction_dct(I) where I is a BW image matrix
Reconstructs the background image by the DCT method proposed in the paper by
L.Chen et al (2008) :
'Automatic TFT-LCD mura defect inspection using discrete cosine transform-based background filtering and ‘just noticeable difference’ quantification strategies'
"""
# Type II DCT is used, the inverse is the Type-III DCT according to SF2 handout
M = I.shape[0] # number of columns in image matrix
N = I.shape[1] # number of rows in image matrix
X = fftpack.dct(fftpack.dct(I, type = 2, norm = 'ortho').T, type = 2, norm = 'ortho').T
X[1:,:][:,1:]= 0;
Y =fftpack.idct( fftpack.idct(X, type = 2, norm = 'ortho').T, type =2, norm = 'ortho').T
Y = np.uint8(Y)
#cv2.imwrite('stage_3.png', Y)
return Y
def inverseATA(self, divBound):
divBound.applyGaussForward()
div = divBound.divergence.div
div = 0.5*fft.dct(div, axis=0)
div = 0.5*fft.dct(div, axis=1)
div = 0.5*fft.dct(div, axis=2)
div[0,0,0] = 0.0
div = div / self.eigvalues
div = fft.idct(div, axis=0) / ( self.M + 1. )
div = fft.idct(div, axis=1) / ( self.N + 1. )
div = fft.idct(div, axis=2) / ( self.P + 1. )
divBound.divergence.div = div
divBound.applyGaussBackward()
return divBound
def _build_test_image(size, edge_width, core):
"""Build a synthetic test image from given DCT coefficients.
The result is a PIL image of size[0] rows and size[1] columns, in the 'F' format. Core is a list
of lists which describes the DCT coefficients to be placed at the core of the image. It must be
rectangular (that is, all lists should have the same length), but not necessarily square. An
initial (size[0]-2*edge_width)x(size[1]-2*edge_width) is generated from these DCT coefficients,
and then embedded in the center of a black image.
The output format is 'F' in order to ensure that Hash.hash_image gets as close to the
coefficients in core as possible.
Hint: ensure that the coefficients in core are not close to a multiple of Hash.dct_coeff_split,
since the rounding used is always floor, and the FP errors might lead to a number such as 1000
turninginto 999.99, which gets rounded to 999 and which is 124 when divided by the standard
Hash.dct_coeff_split of 8, rather than the original 125.
Args:
size: a two-element sequence containing the number of rows and columns of the output image.
edge_width: number of edge pixels in the image.
core: a rectangular list-of-lists containing DCT coefficients used to generate the image.
Returns:
A PIL image as described above.
"""
mat_dct = numpy.zeros((size[0] - 2 * edge_width, size[1] - 2 * edge_width))
row_idx = 0
for row in core:
col_idx = 0
for value in row:
# Yup, it looks wierd. But we need to generate stuff in transpose space so
# Image.formarray picks it up correctly. So we generate the coefficients in a transposd
# form, and they'll be corrected when doing the final transpose.
mat_dct[col_idx, row_idx] = value
col_idx += 1
row_idx += 1
mat_core = fftpack.idct(fftpack.idct(mat_dct, norm='ortho').T, norm='ortho').T
mat = numpy.float32(numpy.random.randint(0, 127, size))
mat[edge_width:size[0] - edge_width, edge_width:size[1] - edge_width] = mat_core
mat += 128
return Image.fromarray(numpy.float32(mat).T, mode='F')
def reconstruct(yuv_tiles, original_shape, q=20):
recon = []
row, col = int(original_shape[0]/8), int(original_shape[1]/8)
if original_shape[0]%8 != 0: row+= 1
if original_shape[1]%8 != 0: col+= 1
for i in range(len(yuv_tiles)):
if i==0: color = 'y'
if i==1: color = 'u'
if i==2: color = 'v'
color_tiles = []
for tile in yuv_tiles[i]:
unquantized_tile = unquantizeTile(tile, color, q)
idct_tile = idct(idct(unquantized_tile.T, norm='ortho').T, norm='ortho')
color_tiles.append(idct_tile)
recon.append(combineTiles(color_tiles, (row, col), 8))
reconstructed = np.zeros((row*8, col*8, 3), dtype=float)
reconstructed[:,:,0], reconstructed[:,:,1], reconstructed[:,:,2] = recon[0], recon[1], recon[2]
cropped = reconstructed[:original_shape[0], :original_shape[1]]
return cropped
def regr(mat):
global qa
new = np.zeros((8,8))
for i in range(8):
for j in range(8):
new[i][j] = mat[i][j]*qa[i][j]
c = idct(new, type=2, n=None, axis=-1, norm='ortho', overwrite_x=False)
for i in range(8):
for j in range(8):
c[i][j] = c[i][j] + 128
return c
def inverse_transform(array):
"""
Args:
list(list(numpy.ndarray)): 2D list of numpy array dct coefficients as float32 of each numpy array
Returns:
list(list(numpy.ndarry)): the idct casted to uint8 of each numpy array in the 2D list
"""
return [
[np.rint(idct(array[x][y], norm='ortho'))
for y in range(0, len(array[x]))]
for x in range(0, len(array))
]
def fbw_kde(data, N=None, MIN=None, MAX=None, overfit_factor=1.0):
# Parameters to set up the mesh on which to calculate
N = 2**14 if N is None else int(2**np.ceil(np.log2(N)))
if MIN is None or MAX is None:
minimum = min(data)
maximum = max(data)
Range = maximum - minimum
MIN = minimum - Range/10 if MIN is None else MIN
MAX = maximum + Range/10 if MAX is None else MAX
# Range of the data
R = MAX-MIN
# Histogram the data to get a crude first approximation of the density
M = len(data)
DataHist, bins = np.histogram(data, bins=N, range=(MIN, MAX))
DataHist = DataHist/M
DCTData = fftpack.dct(DataHist, norm=None)
M = M
I = np.arange(1, N, dtype=np.float64)**2
SqDCTData = np.float64((DCTData[1:]/2.0)**2)
# The fixed point calculation finds the bandwidth = t_star
failure = True
for guess in np.logspace(-1, 2, 20):
try:
t_star = optimize.brentq(fixed_point,
0, guess,
args=(np.float64(M), I, SqDCTData))
failure = False
break
except ValueError:
failure = True
if failure:
raise ValueError('Initial root-finding failed.')
# Smooth the DCTransformed data using t_star divided by an overfitting
# param that allows sub-optimal but allows for "sharper" features
SmDCTData = DCTData*np.exp(-np.arange(N)**2*pisq*t_star/(2*overfit_factor))
# Inverse DCT to get density
density = fftpack.idct(SmDCTData, norm=None)*N/R
mesh = (bins[0:-1]+bins[1:])/2.
bandwidth = np.sqrt(t_star)*R
density = density/np.trapz(density, mesh)
return bandwidth, mesh, density
def process_block(img, weight, quality):
img = img.copy()
# print('process_cube input: {}'.format(img))
this_quality = np.round(np.max(weight)*quality)
if this_quality < 0:
this_quality = 0
if this_quality > quality - 1:
this_quality = quality - 1
img = cv2.cvtColor(img, cv2.COLOR_BGR2LAB)
img = np.float32(img)
img_dct = dct(dct(img, axis=0)/4, axis=1)/4
Q_luma = luminance_tables[:, :, :, this_quality][:, :, 0].astype(np.float32)
Q_chroma = chrominance_tables[:, :, :, this_quality][:, :, 0].astype(np.float32)
img_dct[:, :, 0] /= Q_luma
img_dct[:, :, 1] /= Q_chroma
img_dct[:, :, 2] /= Q_chroma
img_dct = np.round(img_dct)
img_dct[:, :, 0] *= Q_luma
img_dct[:, :, 1] *= Q_chroma
img_dct[:, :, 2] *= Q_chroma
img_processed = idct(idct(img_dct, axis=0)/4, axis=1)/4
img_processed = np.clip(img_processed, 0, 255)
img_processed = np.uint8(img_processed)
img_processed = cv2.cvtColor(img_processed, cv2.COLOR_LAB2BGR)
return img_processed
def zz_idct2(blk):
return idct(idct(blk.T, norm='ortho').T, norm='ortho')
M = np.dot(Mr, Mc)
while (True):
# Capture frame-by-frame
[retval, frame] = cap.read()
cv2.imshow('Original Video, Gruen Komponente', frame[:, :, 1])
#compute magnitude of 2D DCT of green component
#with suitable normalization for the display,
#with norm='ortho' for "energy conservation" in the subbands and for
#invertibiltity without factor:
X = sft.dct(frame[:, :, 1] / 255.0, axis=1, norm='ortho')
X = sft.dct(X, axis=0, norm='ortho')
#Set to zero the 7/8 highest spatial frequencies in each direction:
X = X * M
frame = np.abs(X)
# Display the resulting frame
cv2.imshow('2D-DCT mit Null Setzen der hoechsten Ortsfrequenzen', frame)
#Inverse 2D DCT:
X = sft.idct(X, axis=1, norm='ortho')
x = sft.idct(X, axis=0, norm='ortho')
cv2.imshow('Inverse 2D DCT ohne die hoechsten Ortsfrequenzen', x)
#Keep window open until key 'q' is pressed:
if cv2.waitKey(1) & 0xFF == ord('q'):
break
# When everything done, release the capture
cap.release()
cv2.destroyAllWindows()
def idct2(a):
return idct(idct(a.T, norm='ortho').T, norm='ortho')
def idct3d(x):
x2 = sf.idct(x.copy(), axis=2, norm='ortho')
x1 = sf.idct(x2.copy(), axis=1, norm='ortho')
x0 = sf.idct(x1.copy(), axis=0, norm='ortho')
return x0
def idctn(x, type=2, axes=None, norm="ortho"):
if axes == None: axes = np.arange(len(x.shape))
y = x
for axis in axes:
y = ft.idct(y, type=type, axis=axis, norm="ortho")
return y
def fbwkde(data,
weights=None,
n_dct=None,
min=None,
max=None,
evaluate_dens=True,
evaluate_at=None):
"""Fixed-bandwidth (standard) Gaussian KDE using the Improved
Sheather-Jones bandwidth.
Code adapted for Python from the implementation in Matlab by Zdravko Botev.
Ref: Z. I. Botev, J. F. Grotowski, and D. P. Kroese. Kernel density
estimation via diffusion. The Annals of Statistics, 38(5):2916-2957, 2010.
Parameters
----------
data : array
weights : array or None
n_dct : None or int
Number of points with which to form a regular grid, from `min` to
`max`; histogram values at these points are sent through a discrete
Cosine Transform (DCT), so `n_dct` should be an integer power of 2 for
speed purposes. If None, uses next-highest-power-of-2 above
len(data)*10.
min : float or None
max : float or None
evaluate_dens : bool
evaluate_at : None or array
Returns
-------
bandwidth : float
evaluate_at : array of float
density : None or array of float
"""
if n_dct is None:
n_dct = int(2**np.ceil(np.log2(len(data) * 10)))
assert int(n_dct) == n_dct
n_dct = int(n_dct)
n_datapoints = len(data)
# Parameters to set up the points on which to evaluate the density
if min is None or max is None:
minimum = data.min()
maximum = data.max()
data_range = maximum - minimum
min = minimum - data_range / 2 if min is None else min
max = maximum + data_range / 2 if max is None else max
hist_range = max - min
# Histogram the data to get a crude first approximation of the density
data_hist, bins = np.histogram(data,
bins=n_dct,
range=(min, max),
density=False,
weights=weights)
# Make into a probability mass function
if weights is None:
data_hist = data_hist / n_datapoints
else:
data_hist = data_hist / np.sum(weights)
# Define a minimum bandwidth relative to mean of distances between points
distances = np.diff(np.sort(data))
min_bandwidth = 2 * np.pi * np.mean(distances)
logging.trace('min_bandwidth, 2pi*mean: %.5e', min_bandwidth)
# Solve for the ISJ fixed point to obtain the "optimal" bandwidth
isj_bw, t_star, dct_data = isj_bandwidth(y=data_hist,
n_datapoints=n_datapoints,
x_range=hist_range,
min_bandwidth=min_bandwidth)
# TODO: Detect numerical instability issues here, prior to returning!
if not evaluate_dens:
return isj_bw, evaluate_at, None
if evaluate_at is None:
# Smooth the discrete-cosine-transformed data using t_star
sm_dct_data = dct_data * np.exp(
-np.arange(n_dct)**2 * _PISQ * t_star / 2)
# Inverse DCT to get density
density = fftpack.idct(sm_dct_data, norm=None) * n_dct / hist_range
if np.any(density < 0):
logging.trace(
'ISJ encountered numerical instability in IDCT (resulting in a'
' negative density). Result will be computed without the IDCT.'
)
evaluate_at = (bins[0:-1] + bins[1:]) / 2
else:
evaluate_at = (bins[0:-1] + bins[1:]) / 2
density = density / np.trapz(density, evaluate_at)
return isj_bw, evaluate_at, density
else:
evaluate_at = np.asarray(evaluate_at, dtype=FTYPE)
density = gaussians(x=evaluate_at,
mu=data.astype(FTYPE),
sigma=np.full(shape=n_datapoints,
fill_value=isj_bw,
dtype=FTYPE),
weights=weights)
return isj_bw, evaluate_at, density
def idct2(image_channel):
return fftpack.idct(fftpack.idct(image_channel.T, norm='ortho').T, norm='ortho')
def idct_2d(image):
return fftpack.idct(fftpack.idct(image.T, norm='ortho').T, norm='ortho')
def idct2(a):
return idct(idct(a, axis=0, norm='ortho'), axis=1, norm='ortho')
def idct2d(data):
return fftpack.idct(fftpack.idct(data, norm='ortho').T, norm='ortho').T
# FFT RHS
rhs_p = fft.dct(rhs, type=dct_type, norm=dct_norm)
rhs_p *= 1. / nx
# Construct solution in Fourier space
l = np.arange(0, nx)
cosl = np.cos(np.pi * l / nx)
sol_p = rhs_p / (2. * (cosl - 1.))
# Normalization is directly obtainable in the case of cosine transform (type 2)
sol_p[0] = sol_leveque(0) - np.sum(sol_p[1:])
sol_p[1:] *= 0.5
# Inverse FFT to get real solution
sol = fft.idct(sol_p, type=dct_type, norm=dct_norm)
# Plot solution
fig1 = pl.figure(figsize=(8, 6))
pl.plot(sol, '.', label='DCT solver')
pl.plot(sol_leveque(x), label='analytic')
pl.legend()
fig1.savefig('sol_dct_leveque.png')
if plot_grad_u:
fig2 = pl.figure(figsize=(8, 6))
pl.plot(np.gradient(sol, dx), '.', label='grad DCTsolver')
pl.plot(bc_leveque(x), label='analytic')
pl.plot(np.gradient(sol_leveque(x), dx), '--', label='grad anal. sol.')
pl.legend()
fig2.savefig('bc_dct_leveque.png')
def obj(c00, N, locs, bins, lims):
'''Objective: choose c0 to minimize difference between histograms.'''
c0 = np.zeros(N)
c0[locs] = c00
return dH(Hy, density(idct(c0), bins, lims))
def err_fun(cc, N, bins, lims):
'''Error function for parallel loop.'''
xhat = get_xhat(N, [*cc], bins, lims)
return dH(Hy, density(xhat, bins, lims), mode='l2')
if __name__ == '__main__':
# Assume there is a k-sparse representation,
N = 256 # these choices of N,k give unique solution most of the time
k = 3
cx = np.zeros(N)
idx_true = np.random.choice(np.arange(N), k, False)
cx[idx_true] = 1 # + np.random.normal(0, 1, k)
x = idct(cx)
x /= np.linalg.norm(x)
# We measure a permutation of x
pi_true = np.random.permutation(np.arange(N))
y = x[pi_true]
# Thus the pixel intensity distribution of x is the that of y
lims = (-.5, .5)
Hy, bins = np.histogram(y, bins=N, range=lims)
# Let's try to do things in parallel -- more than twice as fast!
err_fun_partial = partial(err_fun, N=N, bins=bins, lims=lims)
with Pool() as pool:
res = list(
tqdm(pool.imap(err_fun_partial,
def get_2d_idct(coefficients):
# Get 2D Inverse Cosine Transform of Image
return fftpack.idct(fftpack.idct(coefficients.T, norm='ortho').T,
norm='ortho')
import matplotlib.pyplot as plt
plt.close('all')
#%% Use DCT transform from the scipy library
np.set_printoptions(formatter={'float': '{: 0.3f}'.format})
# numpy array
f = np.array([1, 1, 1, 1, 2, 2, 2, 2], dtype='float32')
print("f = ", f)
# apply dct function on array
F = dct(f, norm='ortho')
print("Fu = ", F) #
f_recon = idct(F, norm='ortho')
print("f_recon = ", f_recon)
status = np.allclose(f, f_recon)
print(" f equal f_recon ? ->", status)
#%% Try find the coefficient for F[0], u=0 frequency 0
u = 0
cosv = np.zeros(8)
F = np.zeros(8)
for i in range(8):
if u == 0:
Cu = 1 / np.sqrt(2)
else:
from scipy.fftpack import dct, idct
img = cv2.imread('lenna.jpg', 0)
z = np.zeros([225, 225])
for i in range(28):
for j in range(28):
img1 = img[i * 8:i * 8 + 8, j * 8:j * 8 + 8]
dct0 = dct(img1)
dct1 = np.transpose(dct0)
dct2 = dct(dct1)
for m in range(8):
for n in range(8):
if m >= 4 or n >= 4:
dct2[m][n] = 0
idct0 = idct(dct2)
idct1 = np.transpose(idct0)
idct2 = np.round(idct(idct1) / 256)
z[i * 8:i * 8 + 8, j * 8:j * 8 + 8] = idct2
plt.subplot(121)
plt.imshow(img, cmap='gray')
plt.title('image')
plt.subplot(122)
plt.imshow(z, cmap='gray')
plt.title('compression')
psnr = 10 * np.log10(
(255 * 255) / (1 / (225 * 225) * np.sum(z - img) * np.sum(z - img)))
print('psnr', psnr)
def convert():
original_pic = Image.open("img/1.bmp")
w, h = original_pic.size
d = int(parameter_1.get())
F = int(parameter_2.get())
size = min(w, h)
#FARE CROP
#TAGLIO IMMAGINE IN MODO QUADRATO IN BASE ALLA MISURA MINORE
resto = size % F
if (resto != 0 and w > h):
cropped = original_pic.crop((0, 0, h - resto, h - resto))
cropped.save('img/cropped.bmp')
elif (resto != 0 and h > w):
cropped = original_pic.crop((0, 0, w - resto, w - resto))
cropped.save('img/cropped.bmp')
elif (resto == 0 and w > h):
#immagine rettangolare lato lungo w
cropped = original_pic.crop((0, 0, h, h))
cropped.save('img/cropped.bmp')
else:
taglio = h - w
cropped = original_pic.crop((0, 0, w, w))
cropped.save('img/cropped.bmp')
#CREO BLOCCHI FXF
k = 0 #contatore per nome immagine
w, h = cropped.size
n = w / F
n = int(n)
for y in range(0, n):
for x in range(0, n):
immagineCroppata = cropped.crop(
(x * F, y * F, (x * F) + F, (y * F) + F))
immagineCroppata.save("img/cropped" + str(k) + ".bmp")
k += 1
N = n * n
for k in range(0, N):
original_pic = Image.open("img/cropped" + str(k) + ".bmp")
w1, h1 = original_pic.size
pix_val = list(original_pic.getdata())
data = np.array(pix_val)
shape = (h1, w1)
skatarata = data.reshape(shape)
#DCT
dct1 = dct(skatarata, norm='ortho')
dct2 = dct(dct1.T, norm='ortho')
dct2 = dct2.T
#TAGLIO FREQUENZE D
for i in range(0, h1):
for j in range(0, w1):
if ((i + j >= d)):
dct2[i][j] = 0
#IDCT (DCT INVERSA)
idct1 = idct(dct2, norm='ortho')
idct2 = idct(idct1.T, norm='ortho')
idct2 = idct2.T
#frequenze <0 assegno 0 frequenza >255 assegno 255
for i in range(0, h1):
for j in range(0, w1):
if ((idct2[i][j] < 0)):
idct2[i][j] = 0
if (idct2[i][j] > 255):
idct2[i][j] = 255
else:
round(idct2[i][j])
#converto in tipo idoneo uint8
numbers_matrix = np.array(idct2.astype(np.uint8))
img = Image.fromarray(numbers_matrix)
img.save('img/SKATARAPUMPUM' + str(k) + '.bmp')
#CREO NUOVA IMMAGINE
images = []
for k in range(0, N):
images.append(Image.open('img/SKATARAPUMPUM' + str(k) + '.bmp'))
k = 0
for t in range(0, N, n):
img1 = append_images(images[t:t + n], direction='horizontal')
img1.save("img/PUMPUM" + str(k) + ".bmp")
k += 1
images = []
for k in range(0, n):
images.append(Image.open('img/PUMPUM' + str(k) + '.bmp'))
final = append_images(images, direction='vertical')
final.save("img/FINAL.bmp")
img_elab = final.resize((450, 450))
img_elab = ImageTk.PhotoImage(img_elab)
elab_img.config(image=img_elab)
elab_img.image = img_elab
print('\n直流信号')
print('振幅:', abs_yf[0] / N) # 直流分量的振幅放大了N倍
# 100hz信号
index_100hz = 100 * N // Fs # 波形的频率 = i * Fs / N，倒推计算索引：i = 波形频率 * N / Fs
print('\n100hz波形')
print('振幅:', abs_yf[index_100hz] * 2.0 / N) # 弦波分量的振幅放大了N/2倍
print('相位:', angle_y[index_100hz])
# 150hz信号
index_150hz = 150 * N // Fs # 波形的频率 = i * Fs / N，倒推计算索引：i = 波形频率 * N / Fs
print('\n150hz波形')
print('振幅:', abs_yf[index_150hz] * 2.0 / N) # 弦波分量的振幅放大了N/2倍
print('相位:', angle_y[index_150hz])
print('100hz与150hz相位差:', angle_y[index_150hz] - angle_y[index_100hz])
print('\n')
# DCT变换
import numpy as np
from scipy.fftpack import dct, idct
y = dct(np.array([4., 3., 5., 10., 5., 3.]))
print(y)
import numpy as np
from scipy.fftpack import dct, idct
y = idct(np.array([4., 3., 5., 10., 5., 3.]))
print(y)
def idct_2d(mat):
return fftpack.idct(fftpack.idct(mat.T, norm='ortho').T, norm='ortho')
def _idct2(a):
return idct(idct(a.T, norm="ortho").T, norm="ortho")
def block_idct(x, block_size=8, masked=False, ratio=0.5):
z = np.zeros(shape=x.shape, dtype=np.float32)
num_blocks = int(x.shape[2] / block_size)
mask = np.zeros((x.shape[0], x.shape[1], block_size, block_size))
mask[:, :, :int(block_size * ratio), :int(block_size * ratio)] = 1
for i in range(num_blocks):
for j in range(num_blocks):
submat = x[:, :, (i * block_size):((i + 1) * block_size), (j * block_size):((j + 1) * block_size)]
if masked:
submat = submat * mask
z[:, :, (i * block_size):((i + 1) * block_size), (j * block_size):((j + 1) * block_size)] = idct(idct(submat, axis=3, norm='ortho'), axis=2, norm='ortho')
return z
def idct2(block):
return idct(idct(block.T, norm='ortho').T, norm='ortho')
def baseline_corr_arPLS(self,
lam: float = 1e3,
niter: int = 100,
tol: float = 2e-3,
x0_data=None,
x1_data=None):
"""
Performs baseline correction using asymmetrically reweighted penalized least squares (arPLS). Based on
10.1016/j.csda.2009.09.020 and 10.1039/c4an01061b, utilizes discrete cosine transform to efficiently
perform the calculation.
Correctly smooths only evenly spaced data!!
If x0_data or x1_data is not None, the data range will be used and weight vector will be fixed during fitting.
In those region where data are located, w = 0 and 1 otherwise.
Parameters
----------
lam : float
Lambda - parametrizes the roughness of the smoothed curve.
niter : int
Maximum number of iterations.
tol: float
Tolerance for convergence based on weight matrix.
x0_data : None or int or float
First x value that denotes the signal from data.
x1_data : None or int or float
Last x value that denotes the signal from data.
"""
N = self.data.shape[0]
Lambda = -2 + 2 * np.cos(
np.arange(N) * np.pi /
N) # eigenvalues of 2nd order difference matrix
gamma = 1 / (1 + lam * Lambda * Lambda)
y_orig = self.data[:, 1].copy()
z = y_orig # initialize baseline
y_corr = None # data corrected for baseline
w = np.ones_like(z) # weight vector
i = 0
crit = 1
fix_w = False
if x0_data or x1_data:
start, end = self._get_start_end_indexes(x0_data, x1_data)
w[start:end] = 0 # data have zero weight
fix_w = True
while crit > tol and i < niter:
z = idct(gamma * dct(w * (y_orig - z) + z, norm='ortho'),
norm='ortho') # calculate the baseline
y_corr = y_orig - z # data corrected for baseline
y_corr_neg = y_corr[y_corr < 0] # negative data values
m = np.mean(y_corr_neg)
s = np.std(y_corr_neg)
new_w = 1 / (1 + np.exp(2 * (y_corr - (2 * s - m)) / s)
) # update weights with logistic function
crit = norm(new_w - w) / norm(new_w)
if not fix_w:
w = new_w
if (i + 1) % int(np.sqrt(niter)) == 0:
print(f'Iteration={i + 1}, {crit=:.2g}')
i += 1
self.data[:, 1] = y_corr
return self, Spectrum.from_xy_values(self.data[:, 0], z, f'{self.name} - baseline'),\
Spectrum.from_xy_values(self.data[:, 0], y_orig, f'{self.name} - original data'),\
Spectrum.from_xy_values(self.data[:, 0], w, f'{self.name} - weights') # return the corrected data, baseline and the original data
def dct_window(x, n_coef):
y = dct(x, norm='ortho')
window = np.zeros(x.shape[0])
window[:min(n_coef, len(x))] = 1
yr = idct(y * window, norm='ortho')
return yr[-1]
def idct2(x):
return spfft.idct(spfft.idct(x.T, norm='ortho', axis=0))
def rgb_to_gray(image):
return np.dot(image[..., :3], [0.299, 0.587, 0.144])
Xorig = im.imread('download.jpeg')
print(rgb_to_gray(Xorig).shape)
X = spimg.zoom(rgb_to_gray(Xorig), 0.2)
ny, nx = X.shape
k = round(nx * ny * 0.5) # 50% sample
ri = np.random.choice(nx * ny, k, replace=False) # random sample of indices
b = X.T.flat[ri]
#b = np.expand_dims(b, axis=1)
# create dct matrix operator using kron (memory errors for large ny*nx)
A = np.kron(spfft.idct(np.identity(nx), norm='ortho', axis=0),
spfft.idct(np.identity(ny), norm='ortho', axis=0))
A = A[ri, :] # same as phi times kron
# do L1 optimization
vx = cvx.Variable(nx * ny)
objective = cvx.Minimize(cvx.norm(vx, 1))
constraints = [A * vx == b]
prob = cvx.Problem(objective, constraints)
result = prob.solve(verbose=True)
Xat2 = np.array(vx.value).squeeze()
Xat = Xat2.reshape(nx, ny).T # stack columns
Xa = idct2(Xat)
mask = np.zeros(X.shape)
def idct2(array):
return idct(idct(array, axis=0, norm='ortho'), axis=1, norm='ortho')
def dct_2d_reverse(block):
block = end_T(block)
block = idct(block, norm='ortho')
block = end_T(block)
block = idct(block, norm='ortho')
return block
def convolve(img, gaus_2d, hx, hy):
dat = idct(idct(dct(dct(img, axis=0, norm = 'ortho'), axis=1, \
norm='ortho') * gaus_2d, axis=1, norm='ortho'), axis=0, norm='ortho')[hx, hy]
return dat
print(fftpack.idst(ydst2, 2) / len(x))
print("scipy dst III")
print(fftpack.dst(ydst2, 3) / len(x))
zidst2 = dst_type3(ydst2) / len(x)
print("idst_type2")
print(zidst2)
zmyidst = myidst(ydst2)
print("myidst")
print(zmyidst)
ydst2_ext = np.concatenate([ydst2[1:], [0]])
zmyidst_ext = myidst_ext(ydst2_ext)
print("myidst_ext")
print(zmyidst_ext)
print(fftpack.idct(np.flip(ydst2_ext, 0), 2) / 2)
#expk = 0.5*np.exp(np.arange(N)*1j*np.pi*2/(4*N))
#v = np.zeros_like(expk)
#for k in range(N):
# if k == 0:
# v[k] = expk[k] * (-0 + 1j*ydst2_ext[k])
# else:
# v[k] = expk[k] * (-ydst2_ext[N-k] + 1j*ydst2_ext[k])
#print(np.fft.ifft(v))
pdb.set_trace()