def get_st_d(hsd_in, he_label, he_mode):
DEF = cfg.config_const()
if he_mode == 0:
lab_val = DEF.H_LABEL
if he_mode == 1:
lab_val = DEF.E_LABEL
elif he_mode == 2:
lab_val = DEF.B_LABEL
else:
print("[ERROR]: HE MODE input fail!")
lab_val = -1
np_label = np.array(he_label)
label_mask = (np_label == lab_val)
dat_cnt = np.sum(label_mask)
dat_mask = hsd_in[:, :, 2] * label_mask
val_min = np.min(dat_mask)
val_min_mask = label_mask * val_min
dat_mask_off = dat_mask - val_min_mask
#dat = ()
arr_dat = dat_mask_off.reshape(dat_mask_off.shape[0] *
dat_mask_off.shape[1])
sort_val = sorted(arr_dat)
sort_val.reverse()
outs_val = np.array(sort_val) + val_min
std_val = np.std(outs_val[:dat_cnt])
return std_val
def get_landmask(dat, he_label, cxy_mode, he_mode):
'''
:param dat:[cx,cy,d]
:param he_label: array cx cy d
:param cxy_mode: cx :0 cy:1
:param he_mode: h:0 e:1
:return:
'''
DEF = cfg.config_const()
if he_mode == 0:
lab_val = DEF.H_LABEL
elif he_mode == 1:
lab_val = DEF.E_LABEL
else:
print("[ERROR]: HE MODE input fail!")
lab_val = -1
np_he_label = np.array(he_label)
lab = (np_he_label == lab_val)
dat_cnt = np.sum(lab)
dat_use = dat[:, :, cxy_mode] * lab
dat_min = np.min(dat_use)
dat_off = dat_use - lab * dat_min
array_dat = dat_off.reshape(dat_off.shape[0] * dat_off.shape[1])
array_dat.sort(axis=1) ##?
array_dat = array_dat + dat_min
lm1 = array_dat[int(dat_cnt * 0.99)]
lm25 = array_dat[int(dat_cnt * 0.75)]
lm50 = array_dat[int(dat_cnt * 0.5)]
lm75 = array_dat[int(dat_cnt * 0.25)]
lm99 = array_dat[int(dat_cnt * 0.01)]
return lm1, lm25, lm50, lm75, lm99
def get_feature_mean(hsd_dat, he_label, HE_mode):
'''
:param hsd_dat:hsd
:param he_label: h e
:param mode:0:H 1:E 2:background
:return:
'''
DEF = cfg.config_const()
if HE_mode == 0:
label_val = DEF.H_LABEL
elif HE_mode == 1:
label_val = DEF.E_LABEL
elif HE_mode == 2:
label_val = DEF.B_LABEL
else:
print("[ERROR] the mode input fail!")
label_val = -1
np_label = np.array(he_label)
label_mask = (np_label == label_val)
dat_cnt = np.sum(label_mask)
dat_sum_cx = np.sum(hsd_dat[:, :, 0] * label_mask)
dat_sum_cy = np.sum(hsd_dat[:, :, 1] * label_mask)
dat_sum_d = np.sum(hsd_dat[:, :, 2] * label_mask)
m_cx = dat_sum_cx / dat_cnt
m_cy = dat_sum_cy / dat_cnt
m_d = dat_sum_d / dat_cnt
return [m_cx, m_cy, m_d]
def density_transfer_he(hsd_in, he_label, img_info, tmpl_param, img_st_d,
tmpl_st_d, he_mode):
'''
:param hsd_in:
:param he_label:
:param img_info:
:param tmpl_param:
:param he_mode:
:return:
'''
DEF = cfg.config_const()
if he_mode == 0:
lab_val = DEF.H_LABEL
elif he_mode == 1:
lab_val = DEF.E_LABEL
else:
print("[ERROR]: HE MODE input fail!")
lab_val = -1
hsd_dat = hsd_in * (he_label == lab_val)
hsd_dat_d = hsd_dat[:, :, 2]
tmpl_mean = tmpl_param[0]
tmpl_u = tmpl_mean[lab_val]
img_mean = img_info[0][0]
img_u = img_mean[lab_val]
hsd_dat_d_out = (hsd_dat_d - img_u) / img_st_d * tmpl_st_d + tmpl_u
def main(unused_argv=None):
print_mode()
cfg_run = cfg.config_run()
cfg_const = cfg.config_const()
if cfg.init() != 0:
print("config init fail!")
return 1
if cfg_run.mode == cfg_const.TRAIN_MODE:
train()
elif cfg_run.mode == cfg_const.EVAL_MODE:
eval()
elif cfg_run.mode == cfg_const.TEST_MODE:
test()
else:
print("Input run mode[%s] fail!" % (cfg_run.mode))
return 0
def get_landmask_max_min(dat, cxy_mode, he_label):
'''
:param dat: [cx cy d] data
:param cxy_mode: cx:0 cy 1
:param he_label: h e background label
:return:
'''
DEF = cfg.config_const()
lab = ((he_label == DEF.H_LABEL) | (he_label == DEF.E_LABEL))
dat_cnt = np.sum(lab)
dat_use = dat[:, :, cxy_mode]
dat_min = np.min(dat_use)
dat_off = dat_use - lab * dat_min
arr_dat = dat_off.reshape((dat_off.shape[0] * dat_off.shape[1]))
arr_dat_sorted = sorted(arr_dat)
arr_dat_sorted.reverse()
arr_dat = np.array(arr_dat_sorted) + dat_min
lm_max = arr_dat[0]
lm_min = arr_dat[dat_cnt - 1]
return lm_min, lm_max
def angle_change(hsd_in, angle, he_mode):
'''
:param hsd_in:
:param he_lable:
:param tmpl_angle:
:param he_mode:
:return:
'''
DEF = cfg.config_const()
if he_mode == 0:
lab_val = DEF.H_LABEL
elif he_mode == 1:
lab_val = DEF.E_LABEL
else:
print("[ERROR]: HE MODE input fail!")
lab_val = -1
hsd_dat = hsd_in
hsd_out = hsd_dat * angle[lab_val]
return hsd_out
def background_transfer(hsd_in, he_label, img_info, tmpl_param):
'''
:param hsd_in: hsd_dat[b][:,:,0:2] Cx Cy D of b kinds of pixels
:param he_label:
:param img_info:
:param tmpl_param:
:return: hsd_dat[b][:,:,0:2] Cx Cy D of b kinds of pixels after transform
'''
DEF = cfg.config_const()
img_mean = img_info[0][0]
img_u_b = img_mean[2]
tmpl_mean = tmpl_param[0]
tmpl_u_b = tmpl_mean[2]
hsd_dat = hsd_in * (he_label == DEF.B_LABEL)
hsd_dat_out = hsd_dat - img_u_b + tmpl_u_b
return hsd_dat_out
def landmask_change(hsd_in, he_label, lm_label, img_landmask, tmpl_landmask,
cxy_mode):
'''
:param hsd_in:
:param he_label:
:param lm_label:
:param img_landmask:
:param tmpl_landmask:
:param cxy_mode:
:return:
'''
DEF = cfg.config_const()
hsd_dat = hsd_in[:, :, cxy_mode] * ((he_label == DEF.H_LABEL) |
(he_label == DEF.E_LABEL))
img_scaled = np.zeros(hsd_dat.shape)
for i in range(0, 5):
img_scaled = (img_scaled) + (hsd_dat * (lm_label == i) *
(img_landmask[i] / tmpl_landmask[i]))
return img_scaled
def mean_change(hsd_in, he_label, he_mean, he_mode):
'''
:param hsd_in:
:param he_label:
:param img_info:
:param he_mode:
:return:
'''
DEF = cfg.config_const()
if he_mode == 0:
lab_val = DEF.H_LABEL
elif he_mode == 1:
lab_val = DEF.E_LABEL
else:
print("[ERROR]: HE MODE input fail!")
lab_val = -1
hsd_dat = hsd_in * (he_label == DEF.H_LABEL | he_label == DEF.E_LABEL)
mean = he_mean[lab_val] #h_mean or e_mean
hsd_dat_out = hsd_dat - mean
return hsd_dat_out
def pixels_classify(img, sort_label, labels, w, h):
label_idx = 0
img_shape = img.shape
heb_label = np.zeros((img_shape[0], img_shape[1]))
DEF = cfg.config_const()
for i in range(w):
for j in range(h):
if (labels[label_idx] == sort_label[0]):
heb_label[i][j] = DEF.H_LABEL
if (labels[label_idx] == sort_label[1]):
heb_label[i][j] = DEF.E_LABEL
if (labels[label_idx] == sort_label[2]):
#if a pixel is b classes in k-mean,and its density is lower than 0.2, classify this pixel to background
if (hsl.pixel_rgb2hsd(img[i][j])[2] < 0.2):
heb_label[i][j] = DEF.B_LABEL
# if a pixel is b classes in k-mean,but its density is higher than 0.2, classify this pixel to cytoplasm
# and then change its orig label
else:
heb_label[i][j] = DEF.E_LABEL
label_idx += 1
return heb_label
