TextonsSeg

Repo for PyPi package

Textons - Texture Based Image Segmentation

Textons using LM filters -> Filtros de Leung Malik (https://people.eecs.berkeley.edu/~malik/papers/LM-3dtexton.pdf)

The code here is an implementation of textons using LM filters.

The image is mapped into a 48 dimensions vector space using the features obtained from the LM filters along with the features obtained it also takes into account the RGB values of the pixels

Three more features are added to the end of each vector. These features are row, col and centroids allocated for each pixel, the resulting vector is of the size (1 X 54)

K-Means using Eculidean distance is used to segment image.

inputs:

1. image ==> numpy array
2. number of cluster centers ==> integer
3. number of iterations ==> integers
4. type of colors assignment in the final image ==> integer 0 for 'RANDOM' or 1 for'DEFINED'

output:

• numpy array of image after k means wiht LM filters

Assignment type 'RANDOM' is suited more for higher number of clusters, assignment type 'DEFINED' is better suited when number of clusters is less than 15, however the assignemnt types do not impact the performance of the code in any significant way

Executing the script

Install the TextonsSeg module from PyPi

pip install TextonsSeg==0.0.2

After the install is complete. You can call the module in your python file as shown below:

from TextonsSeg import Textons
import cv2
img = cv2.imread('path to image')
tex = Textons(img, number of centroids, number of iterations, assignment mode)
img = tex.textons()
cv2.imshow('window name', img)
cv2.waitKey(0)
cv2.destroyAllWindows()

Example

from TextonsSeg import Textons
import cv2
im = cv2.imread('image.jpg')
tex = Textons(im, 3, 25, 1)
img = tex.textons()
cv2.imshow('check', img)
cv2.waitKey(0)
cv2.destroyAllWindows()

• Advisable to keep number of iterations high (recommended >= 100) for reproducibility of result

Improved Reproducibility of Result

• Added feature to make the results reproducible

Result with 10 iterations and K value as 3

To-Do

• Add feature to stop k means on the basis of accuracy
• Add minkowski and mahalanobis distance as distance measure