Source code for histo_kit.utils.image

import numpy as np



def gaussian_window(h, w, sigma=None):
    """
    Create a 2D Gaussian weighting window with values normalized to [0, 1].

    The window is symmetric and smoothly attenuates values towards the edges.
    It is used to blend overlapping patches in whole-slide inference
    to avoid visible stitching artifacts.

    Parameters
    ----------
    h : int
        Height of the window.
    w : int
        Width of the window.
    sigma : float, optional
        Standard deviation controlling the spread of the Gaussian.
        If ``None`` (default), the value is set to ``0.5 * max(h, w)``.

    Returns
    -------
    ndarray of shape (h, w)
        A 2D Gaussian window normalized such that its maximum value is 1.

    Notes
    -----
    - The Gaussian is computed independently in x and y dimensions and combined
      via an outer product.
    - Normalization ensures the central peak equals 1.

    Examples
    --------
    >>> win = gaussian_window(5, 5)
    >>> win.shape
    (5, 5)
    >>> win.max()
    1.0
    """
    if sigma is None:
        sigma = 0.5 * max(h, w)

    xv = np.linspace(-1, 1, w)
    yv = np.linspace(-1, 1, h)
    gx = np.exp(-0.5 * (xv / (sigma / max(h,w)))**2)
    gy = np.exp(-0.5 * (yv / (sigma / max(h,w)))**2)
    win = np.outer(gy, gx)

    win = win / win.max()
    return win




def to_tensor_x(x, **kwargs):
    """
    Convert an RGB image from a NumPy array to a PyTorch tensor.

    This function transposes the image from HWC (height, width, channels) format
    to CHW (channels, height, width) format and casts it to `float32`, preparing
    it for input into PyTorch models.

    Parameters
    ----------
    x : ndarray of shape (H, W, C)
        RGB image as a NumPy array.
    **kwargs
        Additional keyword arguments (currently unused).

    Returns
    -------
    tensor : ndarray of dtype float32, shape (C, H, W)
        Transposed and type-cast tensor suitable for PyTorch.

    Notes
    -----
    - The function does not normalize pixel values (e.g., to [0,1]). This should
      be done separately if required by the model.
    - This is a lightweight function and does not create a true `torch.Tensor`.
      To convert to a PyTorch tensor, use `torch.from_numpy(to_tensor_x(x))`.

    Examples
    --------
    >>> x_tensor = to_tensor_x(image_np)
    >>> print(x_tensor.shape)  # (3, H, W)

    .. footbibliography::
    """
    return x.transpose(2, 0, 1).astype('float32')