aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorAaron LI <aaronly.me@outlook.com>2016-10-26 15:33:28 +0800
committerAaron LI <aaronly.me@outlook.com>2016-10-26 15:33:28 +0800
commitc8eecaef2477209db0b4c19fa1811841b798da63 (patch)
treee870f325c873f211b57462c3fbeba2d0f9dc546e
parent9d851a70476667d2790fb5e7a0b25429ac26614e (diff)
downloadfg21sim-c8eecaef2477209db0b4c19fa1811841b798da63.tar.bz2
Add utils/rotate.py: Implement basic image rotation
The "rotate_center()" function rotates a gray-scale image by a given angle about its center, using the area mapping method or simple nearest neighbor. This function is `numba` JIT-optimized, and can replace the `scipy.ndimage.rotate()` function.
-rw-r--r--fg21sim/utils/rotate.py131
1 files changed, 131 insertions, 0 deletions
diff --git a/fg21sim/utils/rotate.py b/fg21sim/utils/rotate.py
new file mode 100644
index 0000000..b89365c
--- /dev/null
+++ b/fg21sim/utils/rotate.py
@@ -0,0 +1,131 @@
+# Copyright (c) 2016 Weitian LI <liweitianux@live.com>
+# MIT license
+
+"""
+Image (only gray-scale image, i.e., matrix) rotate utilities.
+
+References
+----------
+- Leptonica: Rotation
+ http://www.leptonica.com/rotation.html
+- Image rotation by MATLAB without using imrotate
+ https://stackoverflow.com/a/19687481/4856091
+ https://stackoverflow.com/a/19689081/4856091
+"""
+
+
+import numpy as np
+import numba as nb
+
+
+@nb.jit([nb.float64[:, :](nb.int64[:, :], nb.float64, nb.boolean,
+ nb.boolean, nb.float64),
+ nb.float64[:, :](nb.float64[:, :], nb.float64, nb.boolean,
+ nb.boolean, nb.float64)],
+ nopython=True)
+def rotate_center(imgin, angle, interp=True, reshape=True, fill_value=0.0):
+ """Rotate the input image (only gray-scale image currently supported)
+ by a given angle about its center.
+
+ Parameters
+ ----------
+ imgin : 2D `~numpy.ndarray`
+ Input gray-scale image to be rotated
+ angle : float
+ Rotation angle (unit: [ degree ])
+ interp : bool, optional
+ Use the area mapping of the 4 closest input pixels (``interp=True``),
+ which is also the same as "bilinear interpolation",
+ or use the nearest neighbor (``interp=False``) for rotated pixels.
+ reshape : bool, optional
+ Whether adapt the output shape so that the input image is contained
+ completely in the output?
+ fill_value : float, optional
+ Value used to fill pixels in the rotated image that corresponding
+ pixels outside the boundaries of the input image.
+ """
+ nrow, ncol = imgin.shape
+ # Rotation transformation image
+ angle = np.deg2rad(angle)
+ mrotate = np.zeros((2, 2), dtype=np.float64)
+ mrotate[0, 0] = np.cos(angle)
+ mrotate[0, 1] = np.sin(angle)
+ mrotate[1, 0] = -np.sin(angle)
+ mrotate[1, 1] = np.cos(angle)
+ # Determine the shape of rotated image
+ corner00 = np.array((0, 0))
+ corner01 = np.array((0, ncol-1))
+ corner10 = np.array((nrow-1, 0))
+ corner11 = np.array((nrow-1, ncol-1))
+ corners = np.vstack((corner00, corner01, corner10, corner11))
+ if reshape:
+ dest = np.dot(corners.astype(np.float64), mrotate)
+ # XXX: ``numba`` does not support ``np.max()`` with arguments
+ minc = np.min(dest[:, 0])
+ minr = np.min(dest[:, 1])
+ maxc = np.max(dest[:, 0])
+ maxr = np.max(dest[:, 1])
+ nr = int(maxr - minr + 0.5)
+ nc = int(maxc - minc + 0.5)
+ else:
+ # Constraint to be same shape
+ nr = nrow
+ nc = ncol
+ imgout = np.ones((nr, nc)) * fill_value
+ #
+ # Calculate the offset, for easier transformation of rotated pixels
+ # back to input image.
+ #
+ # NOTE:
+ # Notations:
+ # P_out : (r_out, c_out) a pixel in the output rotated image
+ # C_out : center position of the output rotated image
+ # P_in : (r_in, c_in) a pixel in the input image
+ # C_in : center position of the input image
+ # R : rotation matrix
+ # R_T : transposed rotation matrix
+ # The rotation relation between pixel pair is (about the center):
+ # (P_in - C_in) * R = P_out - C_out
+ # Then:
+ # (P_in - C_in) = (P_out - C_out) * R_T
+ # And then:
+ # P_in = C_in + (P_out-C_out) * R_T = P_out*R_T + (C_in - C_out*R_T)
+ # Thus can define the "offset" as:
+ # offset = C_in - C_out * R_T
+ # Then the transformation back to input image is simply given by:
+ # P_in = P_out * R_T + offset
+ #
+ center_in = np.array((nrow/2.0 - 0.5, ncol/2.0 - 0.5))
+ center_out = np.array((nr/2.0 - 0.5, nc/2.0 - 0.5))
+ mrotate_T = mrotate.transpose()
+ offset = center_in - np.dot(center_out, mrotate_T)
+ # Map pixels of the rotated image to the input image
+ for rr in range(nr):
+ for cc in range(nc):
+ p_out = np.array((rr, cc))
+ p_in = np.dot(p_out.astype(np.float64), mrotate_T) + offset
+ if np.all((p_in > corner00) & (p_in < corner11)):
+ # Calculate the pixel value for the rotated image
+ if interp:
+ # Use area mapping of the 4 closest input pixels
+ idx_rf, idx_cf = np.floor(p_in).astype(np.int64)
+ idx_rc, idx_cc = np.ceil(p_in).astype(np.int64)
+ # Calculate the overlapping areas
+ p_r, p_c = p_in
+ p4_area = np.array([(idx_rc - p_r) * (idx_cc - p_c),
+ (idx_rc - p_r) * (p_c - idx_cf),
+ (p_r - idx_rf) * (idx_cc - p_c),
+ (p_r - idx_rf) * (p_c - idx_cf)])
+ p4_val = np.array((imgin[idx_rf, idx_cf],
+ imgin[idx_rf, idx_cc],
+ imgin[idx_rc, idx_cf],
+ imgin[idx_rc, idx_cc]))
+ p_val = np.sum(p4_area * p4_val)
+ else:
+ # Use the nearest neighbor as the rotated value
+ idx_r = round(p_in[0])
+ idx_c = round(p_in[1])
+ p_val = imgin[idx_r, idx_c]
+ #
+ imgout[rr, cc] = p_val
+ return imgout