def get_local_feature_one_vertebra(v, I, label, block_size=block_size):
# step 1: filter the image.
# step 2: extract 1st - 4th order of the filtered image.
# here we use, 6 different scaled gaussian + linear transformation
# (DCT, DST, fft)
slice_num = I.shape[0]
width = I.shape[2]
height = I.shape[1]
feature_dim = get_feature_dim()
feature = np.empty((slice_num, height, width, feature_dim), dtype=np.float)
# first guassian filtered
prev_filtered_dim = 0
for s in sigmas:
print "sigma value: " + str(s)
filtered_dim = get_filtered_dim(s)
for i in range(I.shape[0]):
print "for slice " + str(i)
img = I[i]
f = get_local_feature_one_vertebra_one_slice_one_sigma(img, s)
feature[i, :, :, prev_filtered_dim*moment: (prev_filtered_dim + filtered_dim)*moment]=f
prev_filtered_dim = filtered_dim
# step 3: now that we have feature of all the points in image scale,
# we can extract the feature for all the sample points.
# Based on the paper, IEEE T. on Medical Imaging, 2002, "ASM Segmentation with optimal feature"
# we need to use 5*5 window centered at each landmark.
# note: feature = np.empty((slice_num, height, width, feature_dim),dtype=np.float)
print "start " + str( block_size)+ "x" + str(block_size) + " sampling..."
block_area = block_size*block_size
feature_block = np.empty((8, slice_num*block_area, feature_dim), dtype=np.float)
label_block = np.empty((8, slice_num*block_area), dtype=np.int)
size_1 = block_size/2
size_2 = block_size - size_1
for i in range(8):
idx_x = i*2
idx_y = i*2 + 1
# slice_num * 2k+1 * feature_dim
for n in range(slice_num):
# here block is block centered at landmark
min_x = v[n, idx_x] - size_1
max_x = v[n, idx_x] + size_2
min_y = v[n, idx_y] - size_1
max_y = v[n, idx_y] + size_2
#########################################################
#DEBUG
if n == 0:
from src.util import show_slice
from src.util import read_jpg, crop_slice
pn = 1
fn = '4'
fn = '/Users/ruhansa/Dropbox/spine/data/p' + str(pn) + '/label/'+ fn + '.jpg'
label_jpg = read_jpg(fn)
label_arr = np.array(label_jpg.convert('L'), dtype=np.int16)
x_min = 0
x_max = 512
y_min = 130
y_max = 350
clabel = crop_slice(label_arr, (x_min, y_min), (x_max, y_max)) * 255
clabel[min_x: max_x, min_y: max_y] = 255
show_slice(clabel)
#########################################################
feature_block[i, n*block_area:n*block_area+block_area, :] = feature[n, min_x:max_x, min_y: max_y, : ].reshape(block_area, feature_dim)
label_block[i, n*block_area: n*block_area+block_area] = label[n, min_x:max_x, min_y:max_y].reshape(block_area)
return label_block, feature_block
from src.util import show_volume_dots
dots = np.zeros((10, 2, 8), dtype=np.int16)
for i in range(10):
xs = shapes[i, np.arange(0, 15, 2)].astype(int)
ys = shapes[i, np.arange(1, 16, 2)].astype(int)
dots[i] = np.array((xs, ys))
show_volume_dots(background, dots)
if __name__ == "__main__":
from src.util import read_jpg, crop_slice
pn = 1
fn = "4"
fn = "/Users/ruhansa/Dropbox/spine/data/p" + str(pn) + "/label/" + fn + ".jpg"
label_jpg = read_jpg(fn)
label = np.array(label_jpg.convert("L"), dtype=np.int16)
x_min = 0
x_max = 512
y_min = 130
y_max = 350
clabel = crop_slice(label, (x_min, y_min), (x_max, y_max))
# slice(model.mean, clabel)
volume(shapes, clabel)
# from src.asm_util import show_one_vertebra
# show_one_vertebra(shapes, slice_number=shapes.shape[0],width=220, height=512)