isotropic_erosion

deeptrack.optical.math.isotropic_erosion(mask: ndarray | Tensor, radius: float, *, backend: str = 'numpy', device: device | None = None, dtype: dtype | None = None, channel_axis: int | None = None) → ndarray | Tensor

Apply binary erosion to a mask.

Performs morphological erosion using a structuring element of radius radius. Output is always boolean. Shape is preserved.

• If channel_axis is None:

  • (H, W) → treated as a 2D mask

  • (H, W, Z) → treated as a 3D volume

• If channel_axis is specified:

  • Operation is applied independently for each channel

• Singleton channel:

  • (H, W, 1) is treated as 2D and restored after processing

NumPy backend Uses skimage.morphology.isotropic_erosion, based on Euclidean distance.

Torch backend Uses convolution with a full kernel (square/cubic neighborhood), corresponding to Chebyshev distance. This is not strictly isotropic.

Parameters

masknp.ndarray or torch.Tensor

Input mask. Non-zero values are treated as foreground (mask > 0).

radiusfloat

Radius of the structuring element. If radius <= 0, the input is returned unchanged.

backend{“numpy”, “torch”}, default=”numpy”

Backend used for computation.

devicetorch.device, optional

Device used for torch backend.

dtypetorch.dtype, optional

Data type for torch computations.

channel_axisint or None, optional

Axis corresponding to channels.

Returns

np.ndarray or torch.Tensor

Eroded mask (boolean) with the same shape as the input.