def __init__(self, filename, thresh=10):
self.__fullname = filename
self.__filename = filename[filename.rfind('\\') + 1:]
self.__thresh = thresh
self.__image = cv.LoadImage(filename, cv.CV_LOAD_IMAGE_UNCHANGED)
self.__grayscale = imgtools.normalize_image(cv.LoadImage(filename,
cv.CV_LOAD_IMAGE_GRAYSCALE))
#ext_grayscale = imgtools.image_symmetric_extension(self.__grayscale, max(Surf.SCALES), max(Surf.SCALES))
self.__integralImage = IntegralImage(self.__grayscale)
# max_scale = max(Surf.SCALES)
# self.__integralImage.extend_symmetrically(max_scale, max_scale)
self.__responses = []
self.__features = []
self.__traces = []
self.__orientation = []
self.__descriptors = []
class Surf(object):
OCTAVE_MAP = {9: 1, 15: 1, 21: 1, 27: 2, 39: 2, 51: 3, 75: 3, 99: 4, 147: 4, 195: 5, 291: 5, 387: 5}
SCALES = sorted(OCTAVE_MAP.keys())
FILTER_MAP = np.array([[0, 1, 2, 3],
[1, 3, 4, 5],
[3, 5, 6, 7],
[5, 7, 8, 9],
[7, 9, 10, 11]])
OCTAVES = sorted(OCTAVE_MAP.values())
N = 1
REFINEMENT_ITERATION_COUNT = 5
HESSIAN_RELATIVE_WEIGHT = 0.912
STARTING_ANGLE = 30
LAST_ANGLE = 330
def __init__(self, filename, thresh=10):
self.__fullname = filename
self.__filename = filename[filename.rfind('\\') + 1:]
self.__thresh = thresh
self.__image = cv.LoadImage(filename, cv.CV_LOAD_IMAGE_UNCHANGED)
self.__grayscale = imgtools.normalize_image(cv.LoadImage(filename,
cv.CV_LOAD_IMAGE_GRAYSCALE))
#ext_grayscale = imgtools.image_symmetric_extension(self.__grayscale, max(Surf.SCALES), max(Surf.SCALES))
self.__integralImage = IntegralImage(self.__grayscale)
# max_scale = max(Surf.SCALES)
# self.__integralImage.extend_symmetrically(max_scale, max_scale)
self.__responses = []
self.__features = []
self.__traces = []
self.__orientation = []
self.__descriptors = []
def draw_orientation(self):
image = cv.CloneImage(self.__image)
if self.__features is not None:
for i in range(0, len(self.__features)):
x, y, s, _ = self.__features[i]
size = SIZES[s]
r = size*4
theta = self.__orientation[i]
x, y = size*x, size*y
X, Y = cv.Round(x + r*np.cos(theta)), cv.Round(y + r*np.sin(theta))
cv.Circle(image, (x, y), r, cv.RGB(255, 255, 0), thickness=1)
cv.Line(image, (x, y), (X, Y), cv.RGB(255, 255, 0))
return image
else:
raise NoFeaturesError('No features to draw')
def save_image(self, filename):
cv.SaveImage(filename, self.__image)
def get_features(self):
return self.__features[:]
def get_orientation(self):
return self.__orientation[:]
def get_descriptors(self):
return self.__descriptors[:]
def get_image(self):
return self.__image
def extract_features(self):
print 'Analyzing image', self.__fullname
print ' Computing hessians...'
self.__build_resp_layers()
self.__detect_features()
print ' Finding features...'
self.__find_features()
print ' ', len(self.__features), 'interest points found'
self.__scale_space_refinement()
print ' ', len(self.__features), 'interest points after refinement'
print ' Finding features\' orientation...'
self.__find_orientation()
print ' Building descriptors...'
self.__build_descriptors()
print 'Extraction success!'
self.__save_to_file(self.__fullname + '_points.txt')
self.__grayscale = None
self.__integralImage = None
self.__responses = []
self.__traces = []
def __build_resp_layers(self):
for s in Surf.SCALES:
print ' Computing hessian in scale ' + str(s) + '...'
octave = Surf.OCTAVE_MAP[s]
scale = 2**(octave - 1)
w, h = self.__grayscale.width / scale, self.__grayscale.height / scale
hessian, trace = self.__build_response(w, h, s, scale)
self.__responses.append(hessian)
self.__traces.append(trace)
cv.SaveImage('D:\\test\\hess\\hessian_' + str(s) + '_' + self.__filename + '.jpg', hessian)
def __build_response(self, w, h, s, scale):
L = s / 3
scale_factor_dxx_dyy = 6*L*(2*L - 1)
scale_factor_dxy = 4*L*L
hessian = cv.CreateImage((w, h), cv.IPL_DEPTH_64F, 1)
trace = cv.CreateImage((w, h), cv.IPL_DEPTH_8U, 1)
for (x, y) in [(x, y) for x in range(0, w) for y in range(0, h)]:
X = x * scale
Y = y * scale
dxx = self.__integralImage.dXX(X, Y, L) / scale_factor_dxx_dyy
dyy = self.__integralImage.dYY(X, Y, L) / scale_factor_dxx_dyy
dxy = self.__integralImage.dXY(X, Y, L) / scale_factor_dxy
hessian[y, x] = dxx*dyy - (Surf.HESSIAN_RELATIVE_WEIGHT*dxy)**2
trace[y, x] = 1 if (dxx + dyy) > 0 else 0
return (hessian, trace)
#Checks whether (x, y, s) is far enough from the borders
def __is_not_outlier(self, x, y, s):
if s in range(1, 11):
sigma = SIGMA[s]
dist = cv.Round(10*cv.Sqrt(2)*sigma + 10)
w, h = self.__responses[s].width, self.__responses[s].height
return (x > dist and x < (w-dist) and
y > dist and y < (h-dist))
else:
return False
def __detect_features(self):
resp_layers = self.__responses
for i in range(0, len(Surf.FILTER_MAP)):
for j in range(1, 3):
b = resp_layers[Surf.FILTER_MAP[i, j-1]]
indM = Surf.FILTER_MAP[i, j]
print ' Finding features of scale ' + str(Surf.SCALES[indM]) + '...'
m = resp_layers[indM]
t = resp_layers[Surf.FILTER_MAP[i, j+1]]
for (x, y) in [(x, y) for x in range(1, m.width - 1) for y in range(1, m.height - 1)]:
if np.abs(m[y, x]) <= self.__thresh:
continue
is_max = True
for (dx, dy) in [(dx, dy) for dx in range(-1, 2) for dy in range(-1, 2)]:
if m[y, x] < max([b[y+dy, x+dx], m[y+dy, x+dx], t[y+dy, x+dx]]):
is_max = False
break
if is_max and self.__is_not_outlier(x, y, indM):
self.__features.append((x, y, indM, self.__traces[indM][y, x]))
img = cv.CloneImage(self.__grayscale)
from stitching.ui import cvdrawing
cvdrawing.draw_features(img, self.__features, (1, 0, 0))
cv.SaveImage('D:\\10.jpg', img)
def __is_maximum(self, x, y, b, m, t):
if np.abs(m[y, x]) <= self.__thresh:
return False
for (dx, dy) in [(dx, dy) for dx in range(-1, 2) for dy in range(-1, 2)]:
if m[y, x] < max([b[y+dy, x+dx], m[y+dy, x+dx], t[y+dy, x+dx]]):
return False
return True
#Getting features
def __find_features(self):
resp_layers = self.__responses
for i in range(0, len(Surf.FILTER_MAP)):
for j in range(1, 3):
b = resp_layers[Surf.FILTER_MAP[i, j-1]]
indM = Surf.FILTER_MAP[i, j]
print ' Finding features of scale ' + str(Surf.SCALES[indM]) + '...'
m = resp_layers[indM]
t = resp_layers[Surf.FILTER_MAP[i, j+1]]
octave_b = Surf.OCTAVES[Surf.FILTER_MAP[i, j-1]]
scale_b = 2**(octave_b - 1)
octave_m = Surf.OCTAVES[Surf.FILTER_MAP[i, j-1]]
scale_m = 2**(octave_m - 1)
octave_t = Surf.OCTAVES[Surf.FILTER_MAP[i, j+1]]
scale_t = 2**(octave_t - 1)
sf_b = scale_b / float(scale_m)
sf_t = scale_t / float(scale_m)
candidates = mt.nms_2d(m, m.width, m.height, Surf.N)
for (k, n) in candidates:
if np.abs(m[k, n]) <= self.__thresh:
continue
failed = False
for (x, y) in itertools.product(range(k-1, k+2), range(n-1, n+2)):
if failed:
break
failed = b[int(x/sf_b), int(y/sf_b)] >= m[k, n] or t[int(x/sf_t), int(y/sf_t)] >= m[k, n]
if not failed and self.__is_not_outlier(n, k, indM):
self.__features.append((n, k, indM, self.__traces[indM][k, n]))
def __scale_space_refinement(self):
responses = self.__responses
fake_maxima = []
for i in range(0, len(self.__features)):
(x, y, s, l_sign) = self.__features[i]
k = Surf.REFINEMENT_ITERATION_COUNT
failed = False
while k >= 0 and self.__is_not_outlier(x, y, s) and not failed:
k = k - 1
dx = mt.du_dx(responses, x, y, s)
dy = mt.du_dy(responses, x, y, s)
ds = mt.du_ds(responses, x, y, s)
dxx = mt.du2_dx2(responses, x, y, s)
dyy = mt.du2_dy2(responses, x, y, s)
dxy = mt.du2_dxdy(responses, x, y, s)
dss = mt.du2_ds2(responses, x, y, s)
dxds = mt.du2_dxds(responses, x, y, s)
dyds = mt.du2_dyds(responses, x, y, s)
detH = (2*dxds*dxy*dyds - (dxds**2)*dyy - (dxy**2)*dss +
dxx*dyy*dss - dxx*(dyds**2))
if detH <> 0:
delta_x = -(dx*(dyy*dss - dyds**2) +
dy*(dxds*dyds - dss*dxy) +
ds*(dxy*dyds - dxds*dyy)) / detH
delta_y = -(dx*(dxy*dss - dxds*dyds) +
dy*(dxds**2 - dxx*dss) +
ds*(dxx*dyds - dxds*dxy)) / detH
delta_s = -(dx*(dxy*dyds - dxds*dyy) +
dy*(dxds*dxy - dxx*dyds) +
ds*(dxx*dyy - dxy**2)) / detH
if (np.abs(delta_x) < 1 and np.abs(delta_y) < 1 and
np.abs(delta_s) < 0.4*(2**Surf.OCTAVES[s])):
break
else:
newX = x + delta_x + 0.5
newY = y + delta_y + 0.5
newS = int(s + delta_s/(0.4*(2**Surf.OCTAVES[s])) + 0.5)
if not (newS in range(0, 11)):
failed = True
break
oldSize, newSize = SIZES[s], SIZES[newS]
scale = newSize / float(oldSize)
if newSize > oldSize:
newX, newY = newX / scale, newY / scale
elif newSize < oldSize:
newX, newY = newX * scale, newY * scale
(x, y, s) = (int(newX), int(newY), newS)
else:
break
if (failed or k == -1 or not self.__is_not_outlier(x, y, s) or
np.abs(x - self.__features[i][0]) > Surf.REFINEMENT_ITERATION_COUNT or
np.abs(y - self.__features[i][1]) > Surf.REFINEMENT_ITERATION_COUNT):
fake_maxima.append(self.__features[i])
else:
self.__features[i] = (x, y, s, l_sign)
self.__features = mt.list_subtraction(self.__features, fake_maxima)
def __find_orientation(self):
for (x, y, s, _) in self.__features:
sigma = SIGMA[s]
l = int(sigma)
haar_respX = {}
haar_respY = {}
angle = {}
indexes = set(itertools.product(range(-6, 7), range(-6, 7)))
for (i, j) in [elem for elem in indexes if (elem[0]**2 + elem[1]**2) <= 36]:
X, Y = cv.Round(x + i * sigma), cv.Round(y + j * sigma)
gauss = mt.gaussian(X-x, Y-y, 2*sigma)
haar_respX[(i, j)] = self.__integralImage.haarX(X, Y, l) * gauss
haar_respY[(i, j)] = self.__integralImage.haarY(X, Y, l) * gauss
angle[(i, j)] = cv.FastArctan(haar_respY[(i, j)], haar_respX[(i, j)])
max_length = 0
max_resp_x = max_resp_y = 0
for theta in range(Surf.STARTING_ANGLE, Surf.LAST_ANGLE):
resp_x = resp_y = 0
for p in [index for index in angle.keys() if (theta - 30 <= angle[index] and
angle[index] <= theta + 30)]:
resp_x += haar_respX[p]
resp_y += haar_respY[p]
resp_len = resp_x**2 + resp_y**2
if resp_len > max_length:
max_resp_x, max_resp_y = resp_x, resp_y
max_length = resp_len
self.__orientation.append(cv.FastArctan(max_resp_y, max_resp_x) * mt.TO_RADS_RATIO)
def __build_descriptors(self):
for k in range(0, len(self.__features)):
(x, y, s, _) = self.__features[k]
sigma = SIGMA[s]
l = int(0.5*sigma)
theta = self.__orientation[k]
cosT, sinT = np.cos(theta), np.sin(theta)
descriptor = np.zeros((64))
k = 0
for (i, j) in itertools.product(range(0, 4), range(0, 4)):
sum_resp_x = sum_resp_y = sum_abs_resp_x = sum_abs_resp_y = 0
for (m, n) in itertools.product(range(0, 5), range(0, 5)):
X = cv.Round(x + sigma*(cosT*(5*(i-2) + m) - sinT*(5*(j-2) + n)))
Y = cv.Round(y + sigma*(sinT*(5*(i-2) + m) + cosT*(5*(j-2) + n)))
gauss = mt.gaussian(X-x, Y-y, 3.3*sigma)
resp_x = self.__integralImage.haarX(X, Y, l)
resp_y = self.__integralImage.haarY(X, Y, l)
respX = gauss*(-sinT*resp_x + cosT*resp_y)
respY = gauss*(cosT*resp_x + sinT*resp_y)
sum_resp_x += respX
sum_resp_y += respY
sum_abs_resp_x += np.abs(respX)
sum_abs_resp_y += np.abs(respY)
descriptor[k] = sum_resp_x
descriptor[k+1] = sum_resp_y
descriptor[k+2] = sum_abs_resp_x
descriptor[k+3] = sum_abs_resp_y
k = k + 4
norm = np.linalg.norm(descriptor)
self.__descriptors.append(descriptor / norm)
def __save_to_file(self, filename):
output = open(filename, 'w')
for point in self.__features:
output.write(str(point) + '\n')
output.flush()
output.close()