Question

我想将图像处理从Image J（Fiji）转移到Python。
在图像J中，我将图像分成HSB，然后在B通道上使用Moments Auto-threshold。在Python上，我希望具有阈值t的值，从该值开始进行分割。我找不到关于该主题的任何帮助或代码，所以我在这里。
从论文“保持瞬间阈值：一种新方法，蔡”（here）中，我做到了：

import numpy as np
import cv2
import skimage
from skimage import io

img = io.imread("C:\\Users\\Image.tif")
hsv_img = cv2.cvtColor(filt_img, cv2.COLOR_RGB2HSV)    
H, S, B = cv2.split(img) # spliting the image into HSB

B_histo = skimage.exposure.histogram (B) # making the histogram

pix_sum = B.shape[0]*B.shape[1] # calculating the sum of the pixels

#from the paper, calculating the 4 first odrers m0, m1, m2, m3 to get to p0. The name of the further variables stems from the paper. 

pj = B_histo[0] / pix_sum 
pj_z1 = np.power(B_histo[1], 1) * pj
pj_z2 = np.power(B_histo[1], 2) * pj
pj_z3 = np.power(B_histo[1], 3) * pj

m0 = np.sum(pj)
m1 = np.sum(pj_z1)
m2 = np.sum(pj_z2)
m3 = np.sum(pj_z3)

cd = (m0*m2) - (m1*m1)
c0 = ((-m2*m2) - (-m3*m1))/cd
c1 = ((m0*-m3) - (m1*-m2))/cd

z0 = 0.5 *(-c1 - (np.power(np.power(c1, 2) - 4*c0, 1/2)))
z1 = 0.5 *(-c1 + (np.power(np.power(c1, 2) - 4*c0, 1/2)))

pd = z1 - z0
p0 = (z1 - m1) / pd # p0 should be the percentage of the pixels to which the value 0 is attributed

# using cumulative histogram and comparing it to a target value by calculating the difference. When the difference is the lowest, the index indicates the value of the threshold t
cum_pix = np.cumsum(B_histo[0]) 
target_value = p0 * pix_sum
#td = cum_pix
diff = [abs(i - target_value) for i in cum_pix]
cum_pix[1]
diff[0]
t = [abs(i - target_value) for i in cum_pix].index(np.min(diff))

print(t)

抱歉，代码混乱。无论如何，我在Python上计算的值与在Image J上计算的值不同。您知道问题可能出在哪里吗？还是Python中有一个函数可以检索auto-threshold时刻的值？我会非常感谢您挖掘的技巧或提示，谢谢

Answer 1

矩量和与阈值的比例的计算结果很好，但是skimage直方图存在一个问题，即会缩小bin（即使指定了256 bin）并计算 t 的方式。这是解决问题的方法：

#Instead of skimage.exposure.histogram (B, nbins=256)   
grey_value = np.arange(256)
B_freq = cv2.calcHist([S], [0],None,[256],[0,255])
B_freq = B_freq.reshape((256,))
B_freq = np.int64(B_freq)
B_histo = (B_freq, grey_value)

pix_sum = B.shape[0]*B.shape[1] # calculating the sum of the pixels

#from the paper, calculating the 3 first odrers

pj = B_histo[0] / pix_sum 
pj_z1 = np.power(B_histo[1], 1) * pj
pj_z2 = np.power(B_histo[1], 2) * pj
pj_z3 = np.power(B_histo[1], 3) * pj

m0 = np.sum(pj)
m1 = np.sum(pj_z1)
m2 = np.sum(pj_z2)
m3 = np.sum(pj_z3)

cd = (m0*m2) - (m1*m1)
c0 = ((-m2*m2) - (-m3*m1))/cd
c1 = ((m0*-m3) - (m1*-m2))/cd


z0 = 0.5 *(-c1 - (np.power(np.power(c1, 2) - 4*c0, 1/2)))
z1 = 0.5 *(-c1 + (np.power(np.power(c1, 2) - 4*c0, 1/2)))

pd = z1 - z0
p0 = (z1 - m1) / pd # p0 should be the percentage of the pixels to which the threshold t should be done

# using cumulative histogram and comparing it to a target value by calculating the difference. When the difference is the lowest, the index indicates the value of the threshold t
cum_pix = np.cumsum(B_freq) 
target_value = p0 * pix_sum

diff = [(i - target_value) for i in cum_pix]

def find_nearest(array, value):
    array = np.asarray(array)
    idx = (np.abs(array - value)).argmin()
    return array[idx]

t = diff.index(find_nearest(diff, 0))

print(t)

因此，此代码为您提供了图像J的Moments自动阈值的阈值。

将自动阈值矩方法从Image J转移到Python

1 个答案: