def do_magic(self, img_path, upper, lower):
bound = BoundingBox([upper[0], upper[1], upper[0], lower[1], lower[0], lower[1], lower[0], upper[1]])
f = Fix()
retMsk, retCorr, sumPts, ptsXY = f.register_mask(img_path)
pts = f.pts.flatten()
self_transform = self.avrg - pts
_, indexes = self.nbrs_clf.kneighbors(pts, n_neighbors=5)
indexes = indexes[0]
lbp = LocalBinaryPatternsDescriptor()
hist = HistDescriptor()
p = np.ndarray((4, 2), buffer=bound.data(), dtype=float)
orig_img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1], cv2.imread(img_path, 1))
orig_lbp_desc = lbp.calculate_descriptor(orig_img_part)
orig_hist_desc = hist.calculate_descriptor(orig_img_part)
# draw_points_on_img(img_path, np.ndarray((4, 2), buffer=bound, dtype=float))
for ind in indexes:
pts = np.ndarray((4, 2), buffer=bound.data(), dtype=float).flatten()
p = new_points(pts, self_transform, self.transform_matrix[ind], self.avrg)
p = p.reshape(4, 2)
img_path = os.path.join(DATABASE_LOCATION, '%03d.png' % self.img_ids[ind])
img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1], cv2.imread(img_path, 1))
lbp_desc = lbp.calculate_descriptor(img_part)
hist_desc = hist.calculate_descriptor(img_part)
lbp_diff = np.linalg.norm(orig_lbp_desc - lbp_desc)
hist_diff = np.linalg.norm(orig_hist_desc - hist_desc)
head, filename = os.path.split(img_path)
yield draw_points_on_img(os.path.join(DATABASE_LOCATION, '%03d.png' % self.img_ids[ind]),
p, show=False), lbp_diff, hist_diff, filename
def do_magic_v2(self, img_path, upper, lower, heap_size=3):
bound = BoundingBox([upper[0], upper[1], upper[0], lower[1], lower[0], lower[1], lower[0], upper[1]])
f = Fix()
retMsk, retCorr, sumPts, ptsXY = f.register_mask(img_path)
pts = f.pts.flatten()
self_transform = self.avrg - pts
lbp = LocalBinaryPatternsDescriptor()
hist = HistDescriptor()
p = np.ndarray((4, 2), buffer=bound.data(), dtype=float)
img = cv2.imread(img_path, 1)
orig_img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1], img)
orig_lbp_desc = lbp.calculate_descriptor(orig_img_part)
orig_hist_desc = hist.calculate_descriptor(orig_img_part)
orig_desc = orig_lbp_desc
# select scale
area = bound.area()
# finding closest scale. we sup closest scale - scale with closest area size
scale = min(self.descriptors_db.cfg.sizes, key=lambda x: abs(area - x[0] * x[1]))
center = bound.center()
descriptor_coordinates = (center / scale).astype(np.int)
orig_desc = self.descriptors_db.calculate_one_descriptor(img, scale, descriptor_coordinates[0],
descriptor_coordinates[1])
heap = Heap(heap_size)
for ind, descriptor_value in self.descriptors_db.get_descriptors(scale, descriptor_coordinates):
descriptor_value = np.asarray(descriptor_value)
dist = np.linalg.norm(orig_desc - descriptor_value)
ind = int(ind.split('.')[0])
heap.push((-dist, ind))
for desc_value, ind in heap.data():
pts = np.ndarray((4, 2), buffer=bound.data(), dtype=float).flatten()
trans_ind = self.img_ids.index(ind)
p = new_points(pts, self_transform, self.transform_matrix[trans_ind], self.avrg)
p = p.reshape(4, 2)
img_path = os.path.join(DATABASE_LOCATION, '%03d.png' % self.img_ids[trans_ind])
img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1], cv2.imread(img_path, 1))
lbp_desc = lbp.calculate_descriptor(img_part)
hist_desc = hist.calculate_descriptor(img_part)
lbp_diff = np.linalg.norm(orig_lbp_desc - lbp_desc)
hist_diff = np.linalg.norm(orig_hist_desc - hist_desc)
head, filename = os.path.split(img_path)
yield draw_points_on_img(os.path.join(DATABASE_LOCATION, '%03d.png' % self.img_ids[trans_ind]),
p, show=False), lbp_diff, hist_diff, filename
def calculate_one_descriptor(self, img, size, i, j):
descriptor = self._get_descriptor()
w, h = size
descriptor.w = w
descriptor.h = h
img_part = crop_image(w * i, h * j, w * (i + 1), h * (j + 1), img)
res = descriptor.calculate_descriptor(img_part, mean=img.mean())
return res
def calculate_one_descriptor(self, img, size, i, j):
descriptor = self._get_descriptor()
w, h = size
descriptor.w = w
descriptor.h = h
img_part = crop_image(w * i, h * j, w * (i + 1), h * (j + 1), img)
res = descriptor.calculate_descriptor(img_part, mean=img.mean())
return res
def magic_method(img_path):
transform_matrix, nbrs_clf, img_ids, avrg = load_database(200)
f = Fix()
retMsk, retCorr, sumPts, ptsXY = f.register_mask(img_path)
pts = f.pts.flatten()
self_transform = avrg - pts
_, indexes = nbrs_clf.kneighbors(pts, n_neighbors=5)
indexes = indexes[0]
draw_points_on_img(img_path,
np.ndarray((4, 2), buffer=np.array([64.0, 64, 64, 128, 128, 128, 128, 64]), dtype=float))
lbp = HistDescriptor()
for ind in indexes:
pts = np.ndarray((4, 2), buffer=np.array([64.0, 64, 64, 128, 128, 128, 128, 64]), dtype=float).flatten()
p = new_points(pts, self_transform, transform_matrix[ind], avrg)
p = p.reshape(4, 2)
img_path = os.path.join(DATABASE_LOCATION, '%03d.png' % img_ids[ind])
draw_points_on_img(img_path, p)
img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1], cv2.imread(img_path, 1))
print lbp.calculate_descriptor(img_part)
return
pathImgMask = "%s_mask.png" % img_path
pathImgMasked = "%s_masked.png" % img_path
pathImgOnMask = "%s_onmask.png" % img_path
pathPtsCSV = "%s_pts.csv" % img_path
print f.pts
if retCorr > 0.7:
cv2.imwrite(pathImgMask, f.newMsk)
cv2.imwrite(pathImgMasked, f.newImgMsk)
cv2.imwrite(pathImgOnMask, f.newImgOnMsk)
np.savetxt(pathPtsCSV, f.pts, delimiter=',', newline='\n', fmt='%0.1f')
else:
tmpNewImgMsk = cv2.imread(img_path, 1) # cv2.IMREAD_COLOR)
tmpImgOnMsk = np.zeros((tmpNewImgMsk.shape[0], tmpNewImgMsk.shape[1]), np.uint8)
p00 = (0, 0)
p01 = (0, tmpNewImgMsk.shape[0])
p10 = (tmpNewImgMsk.shape[1], 0)
p11 = (tmpNewImgMsk.shape[1], tmpNewImgMsk.shape[0])
cv2.line(tmpNewImgMsk, p00, p11, (0, 0, 255), 4)
cv2.line(tmpNewImgMsk, p01, p10, (0, 0, 255), 4)
f.newMsk[:] = 0
cv2.imwrite(pathImgMask, f.newMsk)
cv2.imwrite(pathImgMasked, tmpNewImgMsk)
cv2.imwrite(pathImgOnMask, tmpImgOnMsk)
tmpPts = np.zeros((7, 2), np.float64)
np.savetxt(tmpPts, delimiter=',', newline='\n')
fnErr = "%s.err" % img_path
f = open(fnErr, 'w')
f.close()
fzip = "%s.zip" % img_path
zObj = zipfile.ZipFile(fzip, 'w')
zipDir = '%s_dir' % os.path.basename(img_path)
lstFimg = (img_path, pathImgMask, pathImgMasked, pathImgOnMask, pathPtsCSV)
for ff in lstFimg:
ffbn = os.path.basename(ff)
zObj.write(ff, "%s/%s" % (zipDir, ffbn))
print "retCorr = %s" % retCorr
def do_magic(self, img_path, upper, lower):
bound = BoundingBox([
upper[0], upper[1], upper[0], lower[1], lower[0], lower[1],
lower[0], upper[1]
])
f = Fix()
retMsk, retCorr, sumPts, ptsXY = f.register_mask(img_path)
pts = f.pts.flatten()
self_transform = self.avrg - pts
_, indexes = self.nbrs_clf.kneighbors(pts, n_neighbors=5)
indexes = indexes[0]
lbp = LocalBinaryPatternsDescriptor()
hist = HistDescriptor()
p = np.ndarray((4, 2), buffer=bound.data(), dtype=float)
orig_img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1],
cv2.imread(img_path, 1))
orig_lbp_desc = lbp.calculate_descriptor(orig_img_part)
orig_hist_desc = hist.calculate_descriptor(orig_img_part)
# draw_points_on_img(img_path, np.ndarray((4, 2), buffer=bound, dtype=float))
for ind in indexes:
pts = np.ndarray((4, 2), buffer=bound.data(),
dtype=float).flatten()
p = new_points(pts, self_transform, self.transform_matrix[ind],
self.avrg)
p = p.reshape(4, 2)
img_path = os.path.join(DATABASE_LOCATION,
'%03d.png' % self.img_ids[ind])
img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1],
cv2.imread(img_path, 1))
lbp_desc = lbp.calculate_descriptor(img_part)
hist_desc = hist.calculate_descriptor(img_part)
lbp_diff = np.linalg.norm(orig_lbp_desc - lbp_desc)
hist_diff = np.linalg.norm(orig_hist_desc - hist_desc)
head, filename = os.path.split(img_path)
yield draw_points_on_img(os.path.join(
DATABASE_LOCATION, '%03d.png' % self.img_ids[ind]),
p,
show=False), lbp_diff, hist_diff, filename
def calculate_descriptors(self, img_path):
descriptor_template = self._get_descriptor()
img = cv2.imread(img_path, 1)
mean = img.mean()
for w, h in self.cfg.sizes:
descriptor = deepcopy(descriptor_template)
descriptor.w = w
descriptor.h = h
calculated_descriptors = []
for i in range(0, img.shape[0] / w):
for j in range(0, img.shape[1] / h):
img_part = crop_image(w * i, h * j, w * (i + 1), h * (j + 1), img)
res = descriptor.calculate_descriptor(img_part, mean=mean)
calculated_descriptors.append(res)
yield w, h, calculated_descriptors
def run(self):
self.running = True
img = cv2.imread(self.imageFilename, cv2.IMREAD_GRAYSCALE)
if img is None:
self.fail_signal.emit('Failed to load image: {:s}'.format(
self.imageFilename))
img = crop_image(img)
img = cv2.bilateralFilter(img, self.d, self.sigma_color,
self.sigma_space)
Z = img.astype(np.float32).reshape((-1, 1))
criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 100,
1.0)
ret, label, center = cv2.kmeans(Z, 2, None, criteria, 10,
cv2.KMEANS_PP_CENTERS)
center = np.uint8(center)
res = center[label.flatten()]
res2 = res.reshape(img.shape)
# Apply closing then opening.
p3 = np.array(res2, copy=True)
p3[np.where(res2 != min(set(p3.flatten())))] = 255
p3 = cv2.morphologyEx(p3, cv2.MORPH_CLOSE, kernel9)
p3 = cv2.morphologyEx(p3, cv2.MORPH_OPEN, kernel9)
# Apply closing then opening.
p2 = np.array(res2, copy=True)
p2[np.where(p2 != min(set(p2.flatten())))] = 255
p2 = cv2.morphologyEx(p2, cv2.MORPH_OPEN, kernel9)
p2 = cv2.morphologyEx(p2, cv2.MORPH_CLOSE, kernel3)
p2[np.where(p3 != 255)] = 255
seg_img = np.zeros((img.shape[0], img.shape[1], 3), dtype=np.uint8)
seg_img[(p2 == 255) * (p3 == 255)] = colors[0]
seg_img[p2 != 255] = colors[1]
seg_img[p3 != 255] = colors[2]
seg_img = cv2.cvtColor(seg_img, cv2.COLOR_BGR2RGB)
basename = '.'.join(
os.path.basename(self.imageFilename).split('.')[:-1])
self.outputFilename = os.path.join(self.outputPath,
basename + '_segmentation.png')
if self.running:
cv2.imwrite(self.outputFilename, seg_img)
self.succeed_signal.emit()
def calculate_descriptors(self, img_path):
descriptor_template = self._get_descriptor()
img = cv2.imread(img_path, 1)
mean = img.mean()
for w, h in self.cfg.sizes:
descriptor = deepcopy(descriptor_template)
descriptor.w = w
descriptor.h = h
calculated_descriptors = []
for i in range(0, img.shape[0] / w):
for j in range(0, img.shape[1] / h):
img_part = crop_image(w * i, h * j, w * (i + 1),
h * (j + 1), img)
res = descriptor.calculate_descriptor(img_part, mean=mean)
calculated_descriptors.append(res)
yield w, h, calculated_descriptors
def magic_method(img_path):
transform_matrix, nbrs_clf, img_ids, avrg = load_database(200)
f = Fix()
retMsk, retCorr, sumPts, ptsXY = f.register_mask(img_path)
pts = f.pts.flatten()
self_transform = avrg - pts
_, indexes = nbrs_clf.kneighbors(pts, n_neighbors=5)
indexes = indexes[0]
draw_points_on_img(
img_path,
np.ndarray((4, 2),
buffer=np.array([64.0, 64, 64, 128, 128, 128, 128, 64]),
dtype=float))
lbp = HistDescriptor()
for ind in indexes:
pts = np.ndarray(
(4, 2),
buffer=np.array([64.0, 64, 64, 128, 128, 128, 128, 64]),
dtype=float).flatten()
p = new_points(pts, self_transform, transform_matrix[ind], avrg)
p = p.reshape(4, 2)
img_path = os.path.join(DATABASE_LOCATION, '%03d.png' % img_ids[ind])
draw_points_on_img(img_path, p)
img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1],
cv2.imread(img_path, 1))
print lbp.calculate_descriptor(img_part)
return
pathImgMask = "%s_mask.png" % img_path
pathImgMasked = "%s_masked.png" % img_path
pathImgOnMask = "%s_onmask.png" % img_path
pathPtsCSV = "%s_pts.csv" % img_path
print f.pts
if retCorr > 0.7:
cv2.imwrite(pathImgMask, f.newMsk)
cv2.imwrite(pathImgMasked, f.newImgMsk)
cv2.imwrite(pathImgOnMask, f.newImgOnMsk)
np.savetxt(pathPtsCSV, f.pts, delimiter=',', newline='\n', fmt='%0.1f')
else:
tmpNewImgMsk = cv2.imread(img_path, 1) # cv2.IMREAD_COLOR)
tmpImgOnMsk = np.zeros((tmpNewImgMsk.shape[0], tmpNewImgMsk.shape[1]),
np.uint8)
p00 = (0, 0)
p01 = (0, tmpNewImgMsk.shape[0])
p10 = (tmpNewImgMsk.shape[1], 0)
p11 = (tmpNewImgMsk.shape[1], tmpNewImgMsk.shape[0])
cv2.line(tmpNewImgMsk, p00, p11, (0, 0, 255), 4)
cv2.line(tmpNewImgMsk, p01, p10, (0, 0, 255), 4)
f.newMsk[:] = 0
cv2.imwrite(pathImgMask, f.newMsk)
cv2.imwrite(pathImgMasked, tmpNewImgMsk)
cv2.imwrite(pathImgOnMask, tmpImgOnMsk)
tmpPts = np.zeros((7, 2), np.float64)
np.savetxt(tmpPts, delimiter=',', newline='\n')
fnErr = "%s.err" % img_path
f = open(fnErr, 'w')
f.close()
fzip = "%s.zip" % img_path
zObj = zipfile.ZipFile(fzip, 'w')
zipDir = '%s_dir' % os.path.basename(img_path)
lstFimg = (img_path, pathImgMask, pathImgMasked, pathImgOnMask, pathPtsCSV)
for ff in lstFimg:
ffbn = os.path.basename(ff)
zObj.write(ff, "%s/%s" % (zipDir, ffbn))
print "retCorr = %s" % retCorr
def do_magic_v2(self, img_path, upper, lower, heap_size=3):
bound = BoundingBox([
upper[0], upper[1], upper[0], lower[1], lower[0], lower[1],
lower[0], upper[1]
])
f = Fix()
retMsk, retCorr, sumPts, ptsXY = f.register_mask(img_path)
pts = f.pts.flatten()
self_transform = self.avrg - pts
lbp = LocalBinaryPatternsDescriptor()
hist = HistDescriptor()
p = np.ndarray((4, 2), buffer=bound.data(), dtype=float)
img = cv2.imread(img_path, 1)
orig_img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1], img)
orig_lbp_desc = lbp.calculate_descriptor(orig_img_part)
orig_hist_desc = hist.calculate_descriptor(orig_img_part)
orig_desc = orig_lbp_desc
# select scale
area = bound.area()
# finding closest scale. we sup closest scale - scale with closest area size
scale = min(self.descriptors_db.cfg.sizes,
key=lambda x: abs(area - x[0] * x[1]))
center = bound.center()
descriptor_coordinates = (center / scale).astype(np.int)
orig_desc = self.descriptors_db.calculate_one_descriptor(
img, scale, descriptor_coordinates[0], descriptor_coordinates[1])
heap = Heap(heap_size)
for ind, descriptor_value in self.descriptors_db.get_descriptors(
scale, descriptor_coordinates):
descriptor_value = np.asarray(descriptor_value)
dist = np.linalg.norm(orig_desc - descriptor_value)
ind = int(ind.split('.')[0])
heap.push((-dist, ind))
for desc_value, ind in heap.data():
pts = np.ndarray((4, 2), buffer=bound.data(),
dtype=float).flatten()
trans_ind = self.img_ids.index(ind)
p = new_points(pts, self_transform,
self.transform_matrix[trans_ind], self.avrg)
p = p.reshape(4, 2)
img_path = os.path.join(DATABASE_LOCATION,
'%03d.png' % self.img_ids[trans_ind])
img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1],
cv2.imread(img_path, 1))
lbp_desc = lbp.calculate_descriptor(img_part)
hist_desc = hist.calculate_descriptor(img_part)
lbp_diff = np.linalg.norm(orig_lbp_desc - lbp_desc)
hist_diff = np.linalg.norm(orig_hist_desc - hist_desc)
head, filename = os.path.split(img_path)
yield draw_points_on_img(os.path.join(
DATABASE_LOCATION, '%03d.png' % self.img_ids[trans_ind]),
p,
show=False), lbp_diff, hist_diff, filename
def search(descriptors_db, img_template, upper, lower, heap_size=3):
# print upper, lower
bound = BoundingBox(map(float, [upper[0], upper[1], upper[0], lower[1], lower[0], lower[1], lower[0], upper[1]]))
img_path = img_template.format(IMAGE_TO_USE)
lbp = LocalBinaryPatternsDescriptor()
hist = HistDescriptor()
p = np.ndarray((4, 2), buffer=bound.data(), dtype=float)
img = cv2.imread(img_path, 1)
orig_img_part = crop_image(p[0][0], p[0][1], p[2][0], p[2][1], img)
# orig_lbp_desc = lbp.calculate_descriptor(orig_img_part)
# orig_hist_desc = hist.calculate_descriptor(orig_img_part)
area = bound.area()
# finding closest scale. we sup closest scale - scale with closest area size
scale = min(descriptors_db.cfg.sizes, key=lambda x: abs(area - x[0] * x[1]))
center = bound.center()
descriptor_coordinates = (center / scale).astype(np.int)
orig_desc = descriptors_db.calculate_one_descriptor(img, scale, descriptor_coordinates[0],
descriptor_coordinates[1])
heap = Heap(heap_size)
cor_method = 3
_min2 = [100000000000 for i in xrange(4)]
_max2 = [-100000000000 for i in xrange(4)]
# _min = [-0.5799584919619982, -12151325.719406169, 0.011764706578105702, -0.9999839563885284]
# _max = [1.0, -0.0, 4097.904150336981, -0.0]
_min = [-0.7530618036971718, -2386860.0396204507, -2386737.0085002366, -2386737.2483462547]
_max = [1.0593303816872124, 1.0593303816872124, 1509.8481579194934, 1509.8481579194934]
for ind, descriptor_value, (x_pos, y_pos) in descriptors_db.get_descriptors(scale, descriptor_coordinates,
bounding_size=BOUNDING_SIZE):
descriptor_value = np.asarray(descriptor_value)
adding = 0 # sum(orig_desc[6:]) / 2
# dist = scipy.stats.chisquare(f_exp=[x + adding for x in orig_desc[6:]],
# f_obs=[x + adding for x in descriptor_value[6:]])
dist = 0.0
# for i in xrange(3):
# cor_method = i
# _dist = cv2.compareHist(np.float32(orig_desc), np.float32(descriptor_value), OPENCV_METHODS[cor_method][1])
# _dist *= OPENCV_METHODS[cor_method][2]
# dist += _dist / abs((_max[i] - _min[i]))
# if dist < _min2[i]:
# _min2[i] = dist
# if dist > _max2[i]:
# _max2[i] = dist
cor_method = 3
dist = cv2.compareHist(np.float32(orig_desc), np.float32(descriptor_value), OPENCV_METHODS[cor_method][1])
dist *= OPENCV_METHODS[cor_method][2]
try:
dist = hist_distance(orig_desc, descriptor_value) * -1
except:
continue
# dist = sum([- abs(x - y) * abs(x - y) for x, y in zip(descriptor_value, orig_desc)])
# print dist
# raise 1
# print dist
# dist = dist[0]
# abs(sum([w * abs(x - y) for w, (x, y) in zip(weights, zip(orig_desc, descriptor_value))]))
# print '2', dist
ind = ind.split('.')[0]
heap.push((dist, ind, (x_pos, y_pos), descriptor_value))
print _min2, _max2
show_u = tuple(descriptor_coordinates * scale)
show_l = tuple(descriptor_coordinates * scale + scale)
show(IMAGE_TO_USE, show_u, show_l)
for dist, ind, (x_pos, y_pos), descriptor_value in reversed(sorted(heap.data())):
# check descriptor is ok
img = cv2.imread(img_template.format(ind), 1)
descriptor_value2 = descriptors_db.calculate_one_descriptor(img, scale, x_pos, y_pos)
s = sum(descriptor_value2 - descriptor_value)
# if abs(s) > 0.0001:
# raise Exception('wrong descr loaded')
show(ind, (x_pos * scale[0], y_pos * scale[1]), ((x_pos + 1) * scale[0], (y_pos + 1) * scale[1]))
print dist, ind
print ''
def calc(descr, img_path, coords, size=[32, 32]):
img = cv2.imread(img_path, 1)
x, y = coords
x_move, y_move = [_ / 2 for _ in size]
part = crop_image(x - x_move, y - y_move, x + x_move, y + y_move, img)
return descr.calculate_descriptor(part)
