def preprocess(self, tile, ilog=None):
''' Implements some standard preprocessing for all filters. '''
# Get the image tile as a numpy matrix.
if isinstance(tile, pv.Image):
mat = tile.asMatrix2D()
else:
# Assume it is already a matrix
mat = tile
# Apply a logarithmic transformation to pixel values.
if self.norm_log:
mat = np.log(mat + 1)
# print "tile is:",mat
# Transform pixel values to have a mean of zero and unit length
if self.norm_meanunit:
mat = pv.meanUnit(mat)
# print "mat is:",mat
# Window the input tile to reduce fourier edge effects.
window = self._resizeWindow(mat.shape)
if window is not None:
mat = mat * window
if ilog != None:
ilog(pv.Image(mat), label="PREPROCESSED")
return mat
def extract(self, img, face_records):
'''Extract a template that allows the face to be matched.'''
# Compute the 512D vector that describes the face in img identified by
#shape.
im = pv.Image(img[:, :, ::-1])
for face_record in face_records.face_records:
rect = pt.rect_proto2pv(face_record.detection.location)
x, y, w, h = rect.asTuple()
# Extract view
rect = pv.Rect()
cx, cy = x + 0.5 * w, y + 0.5 * h
tmp = 1.5 * max(w, h)
cw, ch = tmp, tmp
crop = pv.AffineFromRect(pv.CenteredRect(cx, cy, cw, ch),
(256, 256))
pvim = pv.Image(img[:, :, ::-1]) # convert rgb to bgr
pvim = crop(pvim)
view = pt.image_pv2proto(pvim)
face_record.view.CopyFrom(view)
tile = pvim.resize((224, 224))
tile = tile.resize((112, 112))
face_im = tile.asOpenCV2()
face_im = face_im[:, :, ::-1] # Convert BGR to RGB
features = self.fr_model.get_embedding(face_im)
face_descriptor = pv.meanUnit(features.flatten())
face_record.template.data.CopyFrom(
pt.vector_np2proto(face_descriptor))
def extractTile(self, im):
import sys
#import numpy as np
from keras.preprocessing import image
#from keras_vggface.vggface import VGGFace
from keras_vggface import utils
assert im.shape == (224, 224, 3)
#tile = tile.resize((224,224))
#tile.show(delay=1000)
img = im
#img = img[:,:,::-1] # Convert BGR to RGB
#mat_ = cv2.cvtColor(mat,cv2.COLOR_RGB2GRAY)
#mat = cv2.cvtColor(mat_,cv2.COLOR_GRAY2RGB)
#img = image.load_img('../image/ajb.jpg', target_size=(224, 224))
img = image.img_to_array(img)
img = np.expand_dims(img, axis=0)
img = utils.preprocess_input(img, version=1) # or version=2
#preds = model.predict(x)
#print('Predicted:', utils.decode_predictions(preds))
#mat = mat[:,:,::-1]
#mat.shape = (1,224,224,3)
# Needed in multithreaded applications
with self.graph.as_default():
tmp = self.recognizer.predict(img)
temp = pv.meanUnit(tmp.flatten())
return temp
def computeVector(self, img):
'''Creates a vector from a face'''
#face = img.asPIL().crop(rect.box()).resize(self.face_size,ANTIALIAS)
vec = img.asMatrix2D().flatten()
if self.norm == PCA_MEAN_STD_NORM:
vec = pv.meanStd(vec)
if self.norm == PCA_MEAN_UNIT_NORM:
vec = pv.meanUnit(vec)
if self.norm == PCA_UNIT_NORM:
vec = pv.unit(vec)
return vec
def computeVector(self,img):
'''Creates a vector from a face'''
#face = img.asPIL().crop(rect.box()).resize(self.face_size,ANTIALIAS)
vec = img.asMatrix2D().flatten()
if self.norm == PCA_MEAN_STD_NORM:
vec = pv.meanStd(vec)
if self.norm == PCA_MEAN_UNIT_NORM:
vec = pv.meanUnit(vec)
if self.norm == PCA_UNIT_NORM:
vec = pv.unit(vec)
return vec
def test_2_meanUnit(self):
'''meanUnit Normalization: norm.mean() = 0.0 and ||norm|| = 1.0....'''
ilog = None
if 'ilog' in list(globals().keys()):
ilog = globals()['ilog']
norm = pv.meanUnit(self.tile)
if ilog != None:
ilog.log(norm,label="meanUnit_Normalization")
mat = norm.asMatrix2D()
self.assertAlmostEqual(mat.mean(),0.0)
length = np.sqrt((mat**2).sum())
self.assertAlmostEqual(length,1.0,places=4)
def test_2_meanUnit(self):
'''meanUnit Normalization: norm.mean() = 0.0 and ||norm|| = 1.0....'''
ilog = None
if 'ilog' in list(globals().keys()):
ilog = globals()['ilog']
norm = pv.meanUnit(self.tile)
if ilog != None:
ilog.log(norm, label="meanUnit_Normalization")
mat = norm.asMatrix2D()
self.assertAlmostEqual(mat.mean(), 0.0)
length = np.sqrt((mat**2).sum())
self.assertAlmostEqual(length, 1.0, places=4)
def PhaseCorrelation(tile1, tile2, phase_only=True, ilog=None):
'''
Uses phase correlation to estimate the best integer displacement to align the images.
Also fits a quadradic to the correltaion surface to determine a sub pixel estimate of
displacement.
Returns four values as a tuple: max_corr, max_displacement, est_corr, est_displacement
max_corr - maximum correlation value.
max_displacement - displacement needed to obtain the maximum correlation. (full pixel)
est_corr - estimated corelation value if subpixel displacement is used.
est_displacement - estimated displacement (subpixel)
see http://en.wikipedia.org/wiki/Phase_correlation
'''
if isinstance(tile1,pv.Image):
tile1 = tile1.asMatrix2D()
else:
tile1 = pv.Image(tile1).asMatrix2D()
raise TypeError("Please pass in a numpy array or a pyvision image.")
if isinstance(tile2,pv.Image):
tile2 = tile2.asMatrix2D()
else:
tile2 = pv.Image(tile2).asMatrix2D()
if tile1.shape != tile2.shape:
raise ValueError("Image tiles must have the same shape. [tile1 %s] != [tile2 %s]"%(tile1.shape,tile2.shape))
# copy the data
tile1 = tile1.copy()
tile2 = tile2.copy()
# normalize the image tiles
tile1 = pv.meanUnit(tile1)
tile2 = pv.meanUnit(tile2)
# compute the fft
Ia = np.fft.fft2(tile1)
Ib = np.conjugate(np.fft.fft2(tile2))
# build the normalized cross-power spectrum
ncs = Ia*Ib
if phase_only:
ncs = ncs/np.abs(ncs)
# build the power spectrum
pc = np.real(np.fft.ifft2(ncs))
if ilog != None:
ilog.log(pv.Image(tile1),label="Tile1")
ilog.log(pv.Image(tile2),label="Tile2")
ilog.log(pv.Image(np.fft.fftshift(pc)),label="Correlation")
max_corr = pc.max()
max_elem = (pc == max_corr).nonzero()
max_elem = max_elem[0][0],max_elem[1][0]
max_point = list(max_elem)
if max_elem[0]*2 > tile1.shape[0]:
max_point[0] = -tile1.shape[0] + max_elem[0]
if max_elem[1]*2 > tile1.shape[1]:
max_point[1] = -tile1.shape[1] + max_elem[1]
est_corr, est_point = QuadradicEstimation(pc,max_point)
return max_corr, max_point, est_corr, est_point
def PhaseCorrelation(tile1, tile2, phase_only=True, ilog=None):
'''
Uses phase correlation to estimate the best integer displacement to align the images.
Also fits a quadradic to the correltaion surface to determine a sub pixel estimate of
displacement.
Returns four values as a tuple: max_corr, max_displacement, est_corr, est_displacement
max_corr - maximum correlation value.
max_displacement - displacement needed to obtain the maximum correlation. (full pixel)
est_corr - estimated corelation value if subpixel displacement is used.
est_displacement - estimated displacement (subpixel)
see http://en.wikipedia.org/wiki/Phase_correlation
'''
if isinstance(tile1, pv.Image):
tile1 = tile1.asMatrix2D()
else:
tile1 = pv.Image(tile1).asMatrix2D()
raise TypeError("Please pass in a numpy array or a pyvision image.")
if isinstance(tile2, pv.Image):
tile2 = tile2.asMatrix2D()
else:
tile2 = pv.Image(tile2).asMatrix2D()
if tile1.shape != tile2.shape:
raise ValueError(
"Image tiles must have the same shape. [tile1 %s] != [tile2 %s]" %
(tile1.shape, tile2.shape))
# copy the data
tile1 = tile1.copy()
tile2 = tile2.copy()
# normalize the image tiles
tile1 = pv.meanUnit(tile1)
tile2 = pv.meanUnit(tile2)
# compute the fft
Ia = np.fft.fft2(tile1)
Ib = np.conjugate(np.fft.fft2(tile2))
# build the normalized cross-power spectrum
ncs = Ia * Ib
if phase_only:
ncs = ncs / np.abs(ncs)
# build the power spectrum
pc = np.real(np.fft.ifft2(ncs))
if ilog != None:
ilog.log(pv.Image(tile1), label="Tile1")
ilog.log(pv.Image(tile2), label="Tile2")
ilog.log(pv.Image(np.fft.fftshift(pc)), label="Correlation")
max_corr = pc.max()
max_elem = (pc == max_corr).nonzero()
max_elem = max_elem[0][0], max_elem[1][0]
max_point = list(max_elem)
if max_elem[0] * 2 > tile1.shape[0]:
max_point[0] = -tile1.shape[0] + max_elem[0]
if max_elem[1] * 2 > tile1.shape[1]:
max_point[1] = -tile1.shape[1] + max_elem[1]
est_corr, est_point = QuadradicEstimation(pc, max_point)
return max_corr, max_point, est_corr, est_point
def extract(self, img, face_records):
'''Extract a template that allows the face to be matched.'''
# Compute the 128D vector that describes the face in img identified by
# shape. In general, if two face descriptor vectors have a Euclidean
# distance between them less than 0.6 then they are from the same
# person, otherwise they are from different people. Here we just print
# the vector to the screen.
im = pv.Image(img[:, :, ::-1])
for face_record in face_records.face_records:
rect = pt.rect_proto2pv(face_record.detection.location)
x, y, w, h = rect.asTuple()
# Extract view
rect = pv.Rect()
cx, cy = x + 0.5 * w, y + 0.5 * h
tmp = 1.5 * max(w, h)
cw, ch = tmp, tmp
crop = pv.AffineFromRect(pv.CenteredRect(cx, cy, cw, ch),
(256, 256))
pvim = pv.Image(img[:, :, ::-1]) # convert rgb to bgr
pvim = crop(pvim)
view = pt.image_pv2proto(pvim)
face_record.view.CopyFrom(view)
# Extract landmarks
l, t, r, b = [int(tmp) for tmp in [x, y, x + w, y + h]]
d = dlib.rectangle(l, t, r, b)
shape = self.shape_pred(img, d)
for i in range(len(shape.parts())):
loc = shape.parts()[i]
landmark = face_record.landmarks.add()
landmark.landmark_id = "point_%02d" % i
landmark.location.x = loc.x
landmark.location.y = loc.y
# Get detection rectangle and crop the face
#rect = pt.rect_proto2pv(face_record.detection.location).rescale(1.5)
#tile = im.crop(rect)
tile = pvim.resize((224, 224))
#tile.show(delay=1000)
face_im = tile.asOpenCV2()
face_im = face_im[:, :, ::-1] # Convert BGR to RGB
#mat_ = cv2.cvtColor(mat,cv2.COLOR_RGB2GRAY)
#mat = cv2.cvtColor(mat_,cv2.COLOR_GRAY2RGB)
#img = image.load_img('../image/ajb.jpg', target_size=(224, 224))
from keras_vggface import utils
from keras.preprocessing import image
face_im = image.img_to_array(face_im)
face_im = np.expand_dims(face_im, axis=0)
face_im = utils.preprocess_input(face_im,
version=2) # or version=2
# Needed in multithreaded applications
with self.graph.as_default():
tmp = self.recognizer.predict(face_im)
face_descriptor = pv.meanUnit(tmp.flatten())
#print('shape:',face_records.face_records[0].landmarks)
#face_descriptor = self.face_rec.compute_face_descriptor(img, shape, JITTER_COUNT)
#face_descriptor = np.array(face_descriptor)
#vec = face_descriptor.flatten()
face_record.template.data.CopyFrom(
pt.vector_np2proto(face_descriptor))