added gaussian low pass filter to image utils

src/cryo_sbi/utils/image_utils.py

@@ -171,6 +171,44 @@ def __call__(self, image: torch.Tensor) -> torch.Tensor:
         return reconstructed
 
 
+class GaussianLowPassFilter:
+    """
+    Low pass filter by dampening the outer frequencies with a Gaussian.
+    """
+
+    def __init__(self, image_size: int, sigma: int):
+        self._image_size = image_size
+        self._sigma = sigma
+        self._grid = torch.linspace(-0.5 * (image_size - 1), 0.5 * (image_size - 1), image_size)
+        self._r_2d = self._grid[None, :] ** 2 + self._grid[:, None] ** 2
+        self._mask = torch.exp(-self._r_2d / (2 * sigma ** 2))
+
+    def __call__(self, image: torch.Tensor) -> torch.Tensor:
+        """
+        Low pass filter an image by dampening the outer frequencies with a Gaussian.
+
+        Args:
+            image (torch.Tensor): Image of shape (n_pixels, n_pixels) or (n_channels, n_pixels, n_pixels).
+
+        Returns:
+            reconstructed (torch.Tensor): Low pass filtered image.
+        """
+
+        fft_image = torch.fft.fft2(image)
+        fft_image = torch.fft.fftshift(fft_image)
+
+        if len(image.shape) == 2:
+            fft_image = fft_image * self._mask
+        elif len(image.shape) == 3:
+            fft_image = fft_image * self._mask.unsqueeze(0)
+        else:
+            raise NotImplementedError
+
+        fft_image = torch.fft.fftshift(fft_image)
+        reconstructed = torch.fft.ifft2(fft_image).real
+        return reconstructed
+
+
 class NormalizeIndividual:
     """
     Normalize an image by subtracting the mean and dividing by the standard deviation.