def texture_features(self, block_size=5, filter_size=3):
"""
Calculates the texture features associated with the image.
block_size gives the size of the texture neighborhood to be processed
filter_size gives the size of the Sobel operator used to find gradient information
"""
#block_size = cv.cvSize(block_size, block_size)
#convert to grayscale float
channels = 1
self.gray_image = cv.cvCreateImage(
cv.cvSize(self.im_width, self.im_height),
cv.IPL_DEPTH_8U, #cv.IPL_DEPTH_16U, #cv.IPL_DEPTH_32F,
channels)
#cv.CV_32FC1, #cv.IPL_DEPTH_32F, #cv.IPL_DEPTH_8U, #cv.IPL_DEPTH_16U,
channels = 1
eig_tex = cv.cvCreateImage(
cv.cvSize(self.im_width * 6, self.im_height), cv.IPL_DEPTH_32F,
channels)
cv.cvCvtColor(self.image, self.gray_image, cv.CV_BGR2GRAY)
#cv.cvAdd(const CvArr* src1, const CvArr* src2, CvArr* dst, const CvArr* mask=NULL );
#highgui.cvConvertImage(self.image, self.gray_image)
cv.cvCornerEigenValsAndVecs(
self.gray_image,
eig_tex, #CvArr* eigenvv,
block_size,
filter_size)
eig_tex = ut.cv2np(eig_tex)
eig_tex = np.reshape(eig_tex, [self.im_height, self.im_width, 6])
#print eig_tex.shape ## [480,640,3]
## (l1, l2, x1, y1, x2, y2), where
## l1, l2 - eigenvalues of M; not sorted
## (x1, y1) - eigenvector corresponding to l1
## (x2, y2) - eigenvector corresponding to l2
tex_feat = np.zeros([3, self.im_height * self.im_width],
dtype=np.float32)
tmp = np.reshape(eig_tex, [self.im_height * self.im_width, 6]).T
s = tmp[0] > tmp[1]
tex_feat[1:3, s] = tmp[0, s] * tmp[2:4, s]
tex_feat[0, s] = tmp[1, s]
tex_feat[1:3, -s] = tmp[1, -s] * tmp[4:6, -s]
tex_feat[0, -s] = tmp[0, -s]
self.tex_feat = tex_feat.T
self.tex_image = np.reshape(self.tex_feat,
[self.im_height, self.im_width, 3])
def texture_features(self, block_size=5, filter_size=3):
"""
Calculates the texture features associated with the image.
block_size gives the size of the texture neighborhood to be processed
filter_size gives the size of the Sobel operator used to find gradient information
"""
#block_size = cv.cvSize(block_size, block_size)
#convert to grayscale float
channels = 1
self.gray_image = cv.cvCreateImage(cv.cvSize(self.im_width, self.im_height),
cv.IPL_DEPTH_8U, #cv.IPL_DEPTH_16U, #cv.IPL_DEPTH_32F,
channels)
#cv.CV_32FC1, #cv.IPL_DEPTH_32F, #cv.IPL_DEPTH_8U, #cv.IPL_DEPTH_16U,
channels = 1
eig_tex = cv.cvCreateImage(cv.cvSize(self.im_width*6, self.im_height),
cv.IPL_DEPTH_32F,
channels)
cv.cvCvtColor(self.image, self.gray_image, cv.CV_BGR2GRAY);
#cv.cvAdd(const CvArr* src1, const CvArr* src2, CvArr* dst, const CvArr* mask=NULL );
#highgui.cvConvertImage(self.image, self.gray_image)
cv.cvCornerEigenValsAndVecs(self.gray_image, eig_tex,#CvArr* eigenvv,
block_size, filter_size)
eig_tex = ut.cv2np(eig_tex)
eig_tex = np.reshape(eig_tex, [self.im_height, self.im_width, 6])
#print eig_tex.shape ## [480,640,3]
## (l1, l2, x1, y1, x2, y2), where
## l1, l2 - eigenvalues of M; not sorted
## (x1, y1) - eigenvector corresponding to l1
## (x2, y2) - eigenvector corresponding to l2
tex_feat = np.zeros([3, self.im_height * self.im_width], dtype=np.float32)
tmp = np.reshape(eig_tex, [self.im_height * self.im_width, 6]).T
s = tmp[0] > tmp[1]
tex_feat[1:3, s] = tmp[0, s] * tmp[2:4, s]
tex_feat[0, s] = tmp[1, s]
tex_feat[1:3, -s] = tmp[1, -s] * tmp[4:6, -s]
tex_feat[0, -s] = tmp[0, -s]
self.tex_feat = tex_feat.T
self.tex_image = np.reshape(self.tex_feat, [self.im_height, self.im_width, 3])
def __init__(self, image, use_texture):
"""
Create an ImageFeatures object for an input image. use_texture is a boolean that when true results in the inclusion of texture features in addition to spatial location and color features.
"""
self.image = image
self.im_size = cv.cvGetSize(image)
self.im_width = self.im_size.width
self.im_height = self.im_size.height
self.im_colors = 3
self.array_image = ut.cv2np(image)
self.tex_feat = None
self.selected_features = None
#self.im_width = image.shape[1]
#self.im_height = image.shape[0]
#self.im_colors = image.shape[2]
self.use_texture = use_texture
self.create_features(self.use_texture)
self.mask_image = None
def __init__(self, image, use_texture):
"""
Create an ImageFeatures object for an input image. use_texture is a boolean that when true results in the inclusion of texture features in addition to spatial location and color features.
"""
self.image = image
self.im_size = cv.cvGetSize(image)
self.im_width = self.im_size.width
self.im_height = self.im_size.height
self.im_colors = 3
self.array_image = ut.cv2np(image)
self.tex_feat = None
self.selected_features = None
#self.im_width = image.shape[1]
#self.im_height = image.shape[0]
#self.im_colors = image.shape[2]
self.use_texture = use_texture
self.create_features(self.use_texture)
self.mask_image = None
def segment_center_object(image,
display_on=False,
nsamp=10000, iter_limit=30,
use_texture=True, use_hsv=True, set_v_to_zero=True,
use_mask=True, remove_saturation=False, remove_boundary = True, prior_gmm=None,
use_flip_heuristic=True):
"""
segment the input image (OpenCV image) and return an ellipse fit to the center object (foreground) and a binary image mask for this foreground object
nsamp : number of pixels to be used when fitting the Gaussian mixture model (impacts speed and accuracy)
iter_limit : maximum number of iterations when fitting the texture model
use_texture : use texture features
use_hsv : use hsv color space for features
set_v_to_zero : effectively remove the value (brightness) component of the hsv features
use_mask : mask out areas of the image prior to training the appearance model and segmenting
remove_saturation : if use_mask, then remove saturated pixels (RGB values = 255 = max value)
remove_boundary : if use_mask, then remove the borders of the image prior to segmenting it
returns a segmentation object (SegmentObject)
"""
#remove_low_freq = True
if use_hsv:
hsv_image = cv.cvCreateImage(cv.cvSize(image.width, image.height),
cv.IPL_DEPTH_8U, 3)
cv.cvCvtColor(image, hsv_image, cv.CV_RGB2HSV) #cv.CV_BGR2HSV)
if set_v_to_zero:
#cvSet(hsv_image, cvScalarAll(0), )
for y in xrange(hsv_image.height):
for x in xrange(hsv_image.width):
pix = hsv_image[y,x]
hsv_image[y,x] = cv.cvScalar(pix[0], pix[1], 0.0)
image = hsv_image
if display_on:
image_list = []
else:
image_list = None
imf = ImageFeatures(image, use_texture=use_texture)
#imf.texture_features()
if use_mask:
#test_mask = np.zeros([image.height, image.width])
#test_mask[0:200, 0:200] = 1.0
#test_mask = test_mask > 0.0
# select saturation mask
nim = ut.cv2np(image)
if remove_saturation:
# remove saturated pixels
#saturation_mask = ~np.alltrue(nim > 255, axis=2)
saturation_mask = ~np.any(nim >= 255, axis=2)
#saturation_mask = np.sum(nim >= 255, axis=2) < 2
if remove_boundary:
# remove boundaries beyond the possible object size
border_y = 50
border_x = 100
too_big_mask = np.zeros(nim.shape[:2], dtype=np.bool)
w = nim.shape[1]
h = nim.shape[0]
too_big_mask[border_y : h - border_y, border_x : w - border_x] = True
if remove_saturation and remove_boundary:
feature_mask = saturation_mask & too_big_mask
elif remove_saturation:
feature_mask = saturation_mask
else:
feature_mask = too_big_mask
disp_mask = feature_mask.copy()
features = imf.select_subset(nsamp, mask_image=feature_mask)
cv_mask = ut.np2cv(disp_mask.astype(np.uint8) * 255)
if image_list is not None:
image_list.append(cv_mask)
else:
features = imf.select_subset(nsamp)
#sego = SegmentObject(image, features, iter_limit=iter_limit)
sego = SegmentObject(image, imf, iter_limit=iter_limit, prior_gmm=prior_gmm)
sego.classify_image(use_flip_heuristic)
sego.clean_classified_image()
#sego.find_largest_object()
sego.find_best_object()
sego.fit_to_largest_object()
if image_list is not None:
image_list.extend(sego.get_images_for_display())
ut.display_images(image_list)
return sego
def segment_center_object(image,
display_on=False,
nsamp=10000,
iter_limit=30,
use_texture=True,
use_hsv=True,
set_v_to_zero=True,
use_mask=True,
remove_saturation=False,
remove_boundary=True,
prior_gmm=None,
use_flip_heuristic=True):
"""
segment the input image (OpenCV image) and return an ellipse fit to the center object (foreground) and a binary image mask for this foreground object
nsamp : number of pixels to be used when fitting the Gaussian mixture model (impacts speed and accuracy)
iter_limit : maximum number of iterations when fitting the texture model
use_texture : use texture features
use_hsv : use hsv color space for features
set_v_to_zero : effectively remove the value (brightness) component of the hsv features
use_mask : mask out areas of the image prior to training the appearance model and segmenting
remove_saturation : if use_mask, then remove saturated pixels (RGB values = 255 = max value)
remove_boundary : if use_mask, then remove the borders of the image prior to segmenting it
returns a segmentation object (SegmentObject)
"""
#remove_low_freq = True
if use_hsv:
hsv_image = cv.cvCreateImage(cv.cvSize(image.width, image.height),
cv.IPL_DEPTH_8U, 3)
cv.cvCvtColor(image, hsv_image, cv.CV_RGB2HSV) #cv.CV_BGR2HSV)
if set_v_to_zero:
#cvSet(hsv_image, cvScalarAll(0), )
for y in xrange(hsv_image.height):
for x in xrange(hsv_image.width):
pix = hsv_image[y, x]
hsv_image[y, x] = cv.cvScalar(pix[0], pix[1], 0.0)
image = hsv_image
if display_on:
image_list = []
else:
image_list = None
imf = ImageFeatures(image, use_texture=use_texture)
#imf.texture_features()
if use_mask:
#test_mask = np.zeros([image.height, image.width])
#test_mask[0:200, 0:200] = 1.0
#test_mask = test_mask > 0.0
# select saturation mask
nim = ut.cv2np(image)
if remove_saturation:
# remove saturated pixels
#saturation_mask = ~np.alltrue(nim > 255, axis=2)
saturation_mask = ~np.any(nim >= 255, axis=2)
#saturation_mask = np.sum(nim >= 255, axis=2) < 2
if remove_boundary:
# remove boundaries beyond the possible object size
border_y = 50
border_x = 100
too_big_mask = np.zeros(nim.shape[:2], dtype=np.bool)
w = nim.shape[1]
h = nim.shape[0]
too_big_mask[border_y:h - border_y, border_x:w - border_x] = True
if remove_saturation and remove_boundary:
feature_mask = saturation_mask & too_big_mask
elif remove_saturation:
feature_mask = saturation_mask
else:
feature_mask = too_big_mask
disp_mask = feature_mask.copy()
features = imf.select_subset(nsamp, mask_image=feature_mask)
cv_mask = ut.np2cv(disp_mask.astype(np.uint8) * 255)
if image_list is not None:
image_list.append(cv_mask)
else:
features = imf.select_subset(nsamp)
#sego = SegmentObject(image, features, iter_limit=iter_limit)
sego = SegmentObject(image,
imf,
iter_limit=iter_limit,
prior_gmm=prior_gmm)
sego.classify_image(use_flip_heuristic)
sego.clean_classified_image()
#sego.find_largest_object()
sego.find_best_object()
sego.fit_to_largest_object()
if image_list is not None:
image_list.extend(sego.get_images_for_display())
ut.display_images(image_list)
return sego
