|
Segmentation procedure
|
Tensor field generation
|
|
Roberts cross is
used on the image to generate a 2x2 matrix
The redundant information makes calculations easier
[Gx*Gx Gx*Gy]
[Gx*Gy Gy*Gy]
|
|
|
| Gaussian Averaging |
|
Image corresponding to each element of the matrix is filtered
independently, then eigenvectors and eigenvalues are extracted.
This results in 3 images:
- dominant direction eigenvalue
- small eigenvalue
- angle of dominant orientation
|
|
|
| Segmentation |
|
This stage depends on the on the actual application, but for the images
that we segmented using sobel operator, followed by thresholding and
non-maximal suppression worked well enough.
|
|
|