Hi, I would like to create a weight map from a binary image in a way to have more intensity values in the border pixels between the close objects. To make it more clear I need to calculate w(x) from an input image following the formula mentioned below. Where w0 and sigma are two constant value, x denotes pixels in the image, d1 denotes the distance to the border of the nearest object and d2 is distance to the border of the second nearest cell.
This is what is done in U-net paper (Article U-Net: Convolutional Networks for Biomedical Image Segmentation
)but I do not how to implement this in python or Matlab
Hello,
To calculate w(x) you need to find all the instances in the ground truth and then compute the distance transform for each individual object. With this distance map, the equation implementation is straightforward as d1(x) and d2(x) are the smallest values in the distance map for pixel x. Please find attached a Matlab implementation of the original equation (including class balance weights w_c) and the result over the input image.
Also you might be interested in this paper: Preprint Multiclass Weighted Loss for Instance Segmentation of Cluttered Cells
Best regards
Hello,
To calculate w(x) you need to find all the instances in the ground truth and then compute the distance transform for each individual object. With this distance map, the equation implementation is straightforward as d1(x) and d2(x) are the smallest values in the distance map for pixel x. Please find attached a Matlab implementation of the original equation (including class balance weights w_c) and the result over the input image.
Also you might be interested in this paper: Preprint Multiclass Weighted Loss for Instance Segmentation of Cluttered Cells
Best regards
Thank you Fidel. It was really helpful
I wonder to know if there is a python implementation for that as well.
Best,
Amirreza
Hello,
You may try this link:
jaidevd.github.io/blog/posts/code/weighted_loss_unet.py
- Badges
- Science topic