pythonnumpyopencvimage-processingimage-morphology
Custom erosion results do not match with OpenCV erosion
I am new to image processing and was trying to write a custom method for erosion and dilation. I then tried to compare my results with OpenCV erosion and dilation function results. I give one padding of zeros to the input image and then overlap the kernel with padded image. Here is my function:
import numpy as np
import matplotlib.pyplot as plt
def operation(image, kernel, padding=0, operation=None):
if operation:
img_operated = image.copy() #this will be the image
"""
The add_padding function below will simply add padding to the image, so the new array with one padding will
look like ->
[[0,0,0,0,0,0,0,0],
[0,0,0,1,1,1,1,0],
[0,0,0,1,1,1,1,0],
[0,1,1,1,1,1,1,0],
[0,1,1,1,1,1,1,0],
[0,1,1,1,1,0,0,0],
[0,1,1,1,1,0,0,0],
[0,0,0,0,0,0,0,0]]
)
"""
image = add_padding(image, padding)
print("Image is \n", image)
print("kernel is \n",kernel)
print("="*40)
vertical_window = padded.shape[0] - kernel.shape[0] #final vertical window position
horizontal_window = padded.shape[1] - kernel.shape[1] #final horizontal window position
print("Vertical Window limit: {}".format(vertical_window))
print("Horizontal Window limit: {}".format(horizontal_window))
print("="*40)
#start with vertical window at 0 position
vertical_pos = 0
values = kernel.flatten() #to compare with values with overlapping element for erosion
#sliding the window vertically
while vertical_pos <= (vertical_window):
horizontal_pos = 0
#sliding the window horizontally
while horizontal_pos <= (horizontal_window):
dilation_flag = False
erosion_flag = False
index_position = 0
#gives the index position of the box
for i in range(vertical_pos, vertical_pos+kernel.shape[0]):
for j in range(horizontal_pos, horizontal_pos+kernel.shape[0]):
#First Case
if operation == "erosion":
if padded[i,j] == values[index_position]:
erosion_flag = True
index_position += 1
else:
erosion_flag = False
break
#Second Case
elif operation == "dilation":
#if we find 1, then break the second loop
if padded[i][j] == 1:
dilation_flag = True
break
else:
return "Operation not understood!"
#if opertion is erosion and there is no match found, break the first 'for' loop
if opr == "erosion" and erosion_flag is False:
break
#if operation is dilation and we find a match, then break the first 'for' loop
if opr == "dilation" and dilation_flag is True:
img_operated[vertical_pos, horizontal_pos] = 1
break
#Check whether erosion flag is true after iterating over one complete overlap
if operation == "erosion" and erosion_flag is True:
img_operated[vertical_pos, horizontal_pos] = 1
elif operation == "erosion" and erosion_flag is False:
img_operated[vertical_pos, horizontal_pos] = 0
#increase the horizontal window position
horizontal_pos += 1
#increase the vertical window position
vertical_pos += 1
return img_operated
return "Operation Required!"
array = np.array([[0,0,1,1,1,1],
[0,0,1,1,1,1],
[1,1,1,1,1,1],
[1,1,1,1,1,1],
[1,1,1,1,0,0],
[1,1,1,1,0,0]], dtype=np.uint8)
kernel = np.array ([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]], dtype = np.uint8)
#image will be padded with one zeros around
result_erosion = operation(array, kernel, 1, "erosion")
result_dilation = operation(array, kernel, 1, "dilation")
#CV2 Erosion and Dilation
cv2_erosion = cv2.erode(array, kernel, iterations=1)
cv2_dilation = cv2.dilate(array, kernel, iterations=1)
The dilation result matches but the erosion result does not. I am not sure why this is the case. Is it because of some padding issues? Does OpenCV pad the image? Or am I implementing the erosion method incorrectly? Here is the image of the results:
Solution
There were two issues with your code:
You weren't checking the value of the kernel. For the dilation this happened to not matter, but you'd see the difference with a different input image.
The erosion was confused. As I mentioned in a comment, the erosion is the complete logical inverse of the dilation. You can think of the erosion as the dilation of the background:
erosion(image) == ~dilation(~image)(with~the logical negation of the image). Therefore, you should be able to use exactly the same code and logic for the erosion as you use for the dilation, but check if you see a background pixel (0) within the kernel, in which case you set that pixel in the output to background (0). To replicate the results of the OpenCV erosion, the padding has to be with foreground (1).
This is the corrected code. I wrote a add_padding function using OpenCV, since it was missing in the OP. The code could be simplified significantly, for example by using a single flag for both operations; by checking the operation string only once at the top of the function and setting a variable with the value 0 or 1 to be used when comparing the input and modifying the output; and by using for loops instead of while loops to iterate over the image. I'll leave those changes to the interested reader.
import numpy as np
import matplotlib.pyplot as plt
import cv2
def add_padding(image, padding, value):
return cv2.copyMakeBorder(image, padding, padding, padding, padding, cv2.BORDER_CONSTANT, value=value)
def operation(image, kernel, padding=0, operation=None):
if operation:
img_operated = image.copy() #this will be the image
padding_value = 0 # <<< ADDED
if operation == "erosion": # <<< ADDED
padding_value = 1 # <<< ADDED
padded = add_padding(image, padding, padding_value) # <<< MODIFIED
vertical_window = padded.shape[0] - kernel.shape[0] #final vertical window position
horizontal_window = padded.shape[1] - kernel.shape[1] #final horizontal window position
#start with vertical window at 0 position
vertical_pos = 0
#sliding the window vertically
while vertical_pos <= vertical_window:
horizontal_pos = 0
#sliding the window horizontally
while horizontal_pos <= horizontal_window:
dilation_flag = False
erosion_flag = False
#gives the index position of the box
for i in range(kernel.shape[0]): # <<< MODIFIED
for j in range(kernel.shape[1]): # <<< MODIFIED
if kernel[i][j] == 1: # <<< ADDED
#First Case
if operation == "erosion":
#if we find 0, then break the second loop
if padded[vertical_pos+i][horizontal_pos+j] == 0: # <<< MODIFIED
erosion_flag = True # <<< MODIFIED
break
#Second Case
elif operation == "dilation":
#if we find 1, then break the second loop
if padded[vertical_pos+i][horizontal_pos+j] == 1: # <<< MODIFIED
dilation_flag = True
break
else:
return "Operation not understood!"
#if opertion is erosion and there is no match found, break the first 'for' loop
if operation == "erosion" and erosion_flag: # <<< MODIFIED
img_operated[vertical_pos, horizontal_pos] = 0 # <<< ADDED
break
#if operation is dilation and we find a match, then break the first 'for' loop
if operation == "dilation" and dilation_flag: # <<< FIXED
img_operated[vertical_pos, horizontal_pos] = 1
break
# !!! Removed unnecessary checks here
#increase the horizontal window position
horizontal_pos += 1
#increase the vertical window position
vertical_pos += 1
return img_operated
return "Operation Required!"
array = np.array([[0,0,1,1,1,1],
[0,0,1,1,1,1],
[1,1,1,1,1,1],
[1,1,1,1,1,1],
[1,1,1,1,0,0],
[1,1,1,1,0,0]], dtype=np.uint8)
kernel = np.array ([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]], dtype = np.uint8)
#image will be padded with one zeros around
result_erosion = operation(array, kernel, 1, "erosion")
result_dilation = operation(array, kernel, 1, "dilation")
#CV2 Erosion and Dilation
cv2_erosion = cv2.erode(array, kernel, iterations=1)
cv2_dilation = cv2.dilate(array, kernel, iterations=1)
- Get the errorbars to bins of a dataset by using bootstrapping
- How to tune PID controller on non-linear model (here: diving of a body)
- django import error - No module named core.management
- Changing tense of text from present/future to past tense
- Discord.py modal form command has no response despite no errors in console
- AttributeError: module 'pyperclip' has no attribute 'waitForPaste'
- Array lists of HTML elements in order by website in selenium / beautiful soup
- Recursive types in Python and difficulties inferring the type of `type(x)(...)`
- `AttributeError` constructing `torchvision.io.VideoReader` in Google Colab
- How to use python selenium to click this element with text bb1
- Base 62 conversion
- How do I get the object if it exists, or None if it does not exist in Django?
- Sampling n= 2000 from a Dask Dataframe of len 18000 generates error Cannot take a larger sample than population when 'replace=False'
- Set same scale in legend matplotlib
- Replicate virtualenv without downloading all the packages again on the same machine
- Tkinter entry widget input are written backwards
- Pydantic v2 vs requests POST JSON - 'not serializable' nightmares
- Unexpected generator behaviour when not assigned to a variable
- How to change the cursor in Pygame to a custom image
- Azure functions in Python - "Duplex option is required" error
- model not showing up in django admin
- How can I see the entire HTTP request that's being sent by my Python application?
- Appending to a list by adding to the previous value in Python
- Layer 'conv2d_11' expected 2 variables, but received 0 variables during loading. Expected: ['conv2d_11/kernel:0', 'conv2d_11/bias:0']
- The Python dbus package is not installed
- How to determine pid of process started via os.system
- Can I store a Parquet file with a dictionary column having mixed types in their values?
- I cannot import monsoon after installing Monsoon
- PyCharm: Why "Python Console" is not accessing ~\.aws\credentials file? How to set it within "Python Console"
- Unexpected Behavior of pd.Grouper with datetime Key and freq Argument