def Bwdist(image, t=8):
if not isinstance(image, np.ndarray):
img = np.array(image)
# print(img.max())
else:
img = image
dist1 = bwdist(np.logical_not(img))
dist2 = bwdist(img)
dist1 = 255 - ((dist1 < t) * 255).astype(np.uint8)
dist2 = 255 - ((dist2 < t) * 255).astype(np.uint8)
b = Image.fromarray(dist1).convert('L')
f = Image.fromarray(dist2).convert('L')
return b, f
def createMask(filtered): from scipy.ndimage import label,binary_erosion from scipy.ndimage import distance_transform_edt as bwdist from skimage.filters import threshold_otsu val = threshold_otsu(filtered); mask = np.zeros(filtered.shape); for z in tqdm.tqdm(range(filtered.shape[2])): label_img = label(filtered[:,:,z]<=val)[0]; mask[:,:,z] = binary_erosion(1-((label_img==label_img[0,0])*1)); dist_mask = bwdist(mask[:,:,0]); radius = np.max(dist_mask); [x,y] = np.where(dist_mask==radius); x=int(x[0]); y=int(y[0]); z = int(filtered.shape[2]/2.0); radius = int(radius/2.0); # obs = filtered[:,:,100:500].copy(); # mobs = mask[:,:,100:500].copy() # mobs = mobs==1; # obs = obs[mobs].flatten() masked = filtered*mask; obs = np.zeros([filtered.shape[0],filtered.shape[1],60]) obs[:,:,:20] = masked[:,:,:20] obs[:,:,20:40] = masked[:,:,z-10:z+10] obs[:,:,40:60] = masked[:,:,filtered.shape[2]-20:filtered.shape[2]] obs = obs[obs>0] return mask,obs;
def cal_wfm(self, pred: np.ndarray, gt: np.ndarray) -> float:
# [Dst,IDXT] = bwdist(dGT);
Dst, Idxt = bwdist(gt == 0, return_indices=True)
# %Pixel dependency
# E = abs(FG-dGT);
E = np.abs(pred - gt)
Et = np.copy(E)
Et[gt == 0] = Et[Idxt[0][gt == 0], Idxt[1][gt == 0]]
# K = fspecial('gaussian',7,5);
# EA = imfilter(Et,K);
K = self.matlab_style_gauss2D((7, 7), sigma=5)
EA = convolve(Et, weights=K, mode="constant", cval=0)
# MIN_E_EA = E;
# MIN_E_EA(GT & EA<E) = EA(GT & EA<E);
MIN_E_EA = np.where(gt & (EA < E), EA, E)
# %Pixel importance
B = np.where(gt == 0, 2 - np.exp(np.log(0.5) / 5 * Dst),
np.ones_like(gt))
Ew = MIN_E_EA * B
TPw = np.sum(gt) - np.sum(Ew[gt == 1])
FPw = np.sum(Ew[gt == 0])
R = 1 - np.mean(Ew[gt == 1])
P = TPw / (TPw + FPw + _EPS)
# % Q = (1+Beta^2)*(R*P)./(eps+R+(Beta.*P));
Q = (1 + self.beta) * R * P / (R + self.beta * P + _EPS)
return Q
def Bwdist(image, t=8):
shape = image.shape[2:]
batch_size = image.shape[0]
image = image.reshape((-1, shape[0], shape[1], 1))
# image = image.squeeze()
image2 = image.copy()
dist1 = image.copy()
dist2 = image.copy()
for i in range(batch_size):
image[i,:,:,:] = bwdist(np.logical_not(image[i,:,:,:]))
image2[i,:,:,:] = bwdist((image2[i,:,:,:]))
dist1[i,:,:,:] = 1.0 - ((image[i,:,:,:]<t)*1.0)
dist2[i,:,:,:] = 1.0 - ((image2[i,:,:,:]<t)*1.0)
dist1 = dist1.reshape((-1, 1, shape[0], shape[1]))
dist2 = dist2.reshape((-1, 1, shape[0], shape[1]))
return dist1, dist2
def cal_wfm(self, pred: np.ndarray, gt: np.ndarray) -> float:
"""
Calculate the weighted F-measure.
"""
# [Dst,IDXT] = bwdist(dGT);
Dst, Idxt = bwdist(gt == 0, return_indices=True)
# %Pixel dependency
# E = abs(FG-dGT);
E = np.abs(pred - gt)
# Et = E;
# Et(~GT)=Et(IDXT(~GT)); %To deal correctly with the edges of the foreground region
Et = np.copy(E)
Et[gt == 0] = Et[Idxt[0][gt == 0], Idxt[1][gt == 0]]
# K = fspecial('gaussian',7,5);
# EA = imfilter(Et,K);
K = self.matlab_style_gauss2D((7, 7), sigma=5)
EA = convolve(Et, weights=K, mode="constant", cval=0)
# MIN_E_EA = E;
# MIN_E_EA(GT & EA<E) = EA(GT & EA<E);
MIN_E_EA = np.where(gt & (EA < E), EA, E)
# %Pixel importance
# B = ones(size(GT));
# B(~GT) = 2-1*exp(log(1-0.5)/5.*Dst(~GT));
# Ew = MIN_E_EA.*B;
B = np.where(gt == 0, 2 - np.exp(np.log(0.5) / 5 * Dst),
np.ones_like(gt))
Ew = MIN_E_EA * B
# TPw = sum(dGT(:)) - sum(sum(Ew(GT)));
# FPw = sum(sum(Ew(~GT)));
TPw = np.sum(gt) - np.sum(Ew[gt == 1])
FPw = np.sum(Ew[gt == 0])
# R = 1- mean2(Ew(GT)); %Weighed Recall
# P = TPw./(eps+TPw+FPw); %Weighted Precision
# 注意这里使用mask索引矩阵的时候不可使用Ew[gt]，这实际上仅在索引Ew的0维度
R = 1 - np.mean(Ew[gt == 1])
P = TPw / (TPw + FPw + _EPS)
# % Q = (1+Beta^2)*(R*P)./(eps+R+(Beta.*P));
Q = (1 + self.beta) * R * P / (R + self.beta * P + _EPS)
return Q