changeset 5:8d1c37190122

Ajouté code de déformations élastiques locales, adapté depuis un travail que j'ai fait la session dernière
author fsavard <francois.savard@polymtl.ca>
date Tue, 26 Jan 2010 14:21:40 -0500
parents 7fb10ae0efec
children 424b5b0d9fcb
files transformations/local_elastic_distortions.py
diffstat 1 files changed, 233 insertions(+), 0 deletions(-) [+]
line wrap: on
line diff
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/transformations/local_elastic_distortions.py	Tue Jan 26 14:21:40 2010 -0500
@@ -0,0 +1,233 @@
+#!/usr/bin/python
+# coding: utf-8
+
+'''
+Implementation of elastic distortions as described in
+Simard, Steinkraus, Platt, "Best Practices for Convolutional
+    Neural Networks Applied to Visual Document Analysis", 2003
+
+Author: François Savard
+Date: Fall 2009, revised Winter 2010
+
+Usage: create the Distorter with proper alpha, sigma etc.
+    Then each time you want to change the distortion field applied,
+    call regenerate_field(). 
+
+    (The point behind this is that regeneration takes some time,
+    so we better reuse the fields a few times)
+'''
+
+import sys
+import math
+import numpy
+import numpy.random
+import scipy.signal # convolve2d
+
+def raw_zeros(size):
+    return [[0 for i in range(size[1])] for j in range(size[0])]
+
+class LocalElasticDistorter():
+    def __init__(self, image_size, kernel_size, sigma, alpha):
+        self.image_size = image_size
+        self.kernel_size = kernel_size
+        self.sigma = sigma
+        self.alpha = alpha
+        self.c_alpha = int(math.ceil(alpha))
+
+        self.kernel = self.gen_gaussian_kernel()
+        self.fields = None
+        self.regenerate_fields()
+
+    # adapted from http://blenderartists.org/forum/showthread.php?t=163361
+    def gen_gaussian_kernel(self):
+        h,w = self.kernel_size
+        a,b = h/2.0, w/2.0
+        y,x = numpy.ogrid[0:w, 0:h]
+        s = self.sigma
+        gauss = numpy.exp(-numpy.square((x-a)/s))*numpy.exp(-numpy.square((y-b)/s))
+        # Normalize so we don't reduce image intensity
+        return gauss/gauss.sum()
+
+    def gen_distortion_field(self):
+        field = numpy.random.uniform(-1.0, 1.0, self.image_size)
+        return scipy.signal.convolve2d(field, self.kernel, mode='same')
+
+    def regenerate_fields(self):
+        '''
+        Here's how the code works:
+        - We first generate "distortion fields" for x and y with these steps:
+            - Uniform noise over [-1, 1] in a matrix of size (h,w)
+            - Blur with a Gaussian kernel of spread sigma
+            - Multiply by alpha
+        - Then (conceptually) to compose the distorted image, we loop over each pixel
+            of the new image and use the corresponding x and y distortions
+            (from the matrices generated above) to identify pixels
+            of the old image from which we fetch color data. As the
+            coordinates are not integer, we interpolate between the
+            4 nearby pixels (top left, top right etc.).
+        - That's just conceptually. Here I'm using matrix operations
+            to speed up the computation. I first identify the 4 nearby
+            pixels in the old image for each pixel in the distorted image.
+            I can then use them as "fancy indices" to extract the proper
+            pixels for each new pixel.
+        - Then I multiply those extracted nearby points by precomputed
+            ratios for the bilinear interpolation.
+        '''
+
+        self.fields = [None, None]
+        self.fields[0] = self.alpha*self.gen_distortion_field()
+        self.fields[1] = self.alpha*self.gen_distortion_field()
+
+        #import pylab
+        #pylab.imshow(self.fields[0])
+        #pylab.show()
+
+        # regenerate distortion index matrices
+        # "_rows" are row indices
+        # "_cols" are column indices
+        # (separated due to the way fancy indexing works in numpy)
+        h,w = self.image_size
+
+        self.matrix_tl_corners_rows = raw_zeros((h,w))
+        self.matrix_tl_corners_cols = raw_zeros((h,w))
+
+        self.matrix_tr_corners_rows = raw_zeros((h,w))
+        self.matrix_tr_corners_cols = raw_zeros((h,w))
+
+        self.matrix_bl_corners_rows = raw_zeros((h,w))
+        self.matrix_bl_corners_cols = raw_zeros((h,w))
+
+        self.matrix_br_corners_rows = raw_zeros((h,w))
+        self.matrix_br_corners_cols = raw_zeros((h,w))
+
+        # those will hold the precomputed ratios for
+        # bilinear interpolation
+        self.matrix_tl_multiply = numpy.zeros((h,w))
+        self.matrix_tr_multiply = numpy.zeros((h,w))
+        self.matrix_bl_multiply = numpy.zeros((h,w))
+        self.matrix_br_multiply = numpy.zeros((h,w))
+
+        for y in range(h):
+            for x in range(w):
+                distort_x = self.fields[0][y,x]
+                distort_y = self.fields[1][y,x]
+                f_dy = int(math.floor(distort_y))
+                f_dx = int(math.floor(distort_x))
+                y0 = y+f_dy
+                x0 = x+f_dx
+                index_tl = [y0, x0]
+                index_tr = [y0, x0+1]
+                index_bl = [y0+1, x0]
+                index_br = [y0+1, x0+1]
+                x_ratio = abs(distort_x-f_dx) # ratio of left vs right (for bilinear)
+                y_ratio = abs(distort_y-f_dy) # ratio of top vs bottom
+
+                # We use a default background color of 0 for displacements
+                # outside of boundaries of the image.
+
+                # if top left outside bounds
+                if index_tl[0] < 0 or index_tl[0] >= h or index_tl[1] < 0 or index_tl[1] >= w: 
+                    self.matrix_tl_corners_rows[y][x] = 0
+                    self.matrix_tl_corners_cols[y][x] = 0
+                    self.matrix_tl_multiply[y,x] = 0
+                else:
+                    self.matrix_tl_corners_rows[y][x] = index_tl[0]
+                    self.matrix_tl_corners_cols[y][x] = index_tl[1]
+                    self.matrix_tl_multiply[y,x] = x_ratio*y_ratio
+
+
+                # if top right outside bounds
+                if index_tr[0] < 0 or index_tr[0] >= h or index_tr[1] < 0 or index_tr[1] >= w:
+                    self.matrix_tr_corners_rows[y][x] = 0
+                    self.matrix_tr_corners_cols[y][x] = 0
+                    self.matrix_tr_multiply[y,x] = 0
+                else:
+                    self.matrix_tr_corners_rows[y][x] = index_tr[0]
+                    self.matrix_tr_corners_cols[y][x] = index_tr[1]
+                    self.matrix_tr_multiply[y,x] = (1.0-x_ratio)*y_ratio
+
+                # if bottom left outside bounds
+                if index_bl[0] < 0 or index_bl[0] >= h or index_bl[1] < 0 or index_bl[1] >= w:
+                    self.matrix_bl_corners_rows[y][x] = 0
+                    self.matrix_bl_corners_cols[y][x] = 0
+                    self.matrix_bl_multiply[y,x] = 0
+                else:
+                    self.matrix_bl_corners_rows[y][x] = index_bl[0]
+                    self.matrix_bl_corners_cols[y][x] = index_bl[1]
+                    self.matrix_bl_multiply[y,x] = x_ratio*(1.0-y_ratio)
+
+                # if bottom right outside bounds
+                if index_br[0] < 0 or index_br[0] >= h or index_br[1] < 0 or index_br[1] >= w:
+                    self.matrix_br_corners_rows[y][x] = 0
+                    self.matrix_br_corners_cols[y][x] = 0
+                    self.matrix_br_multiply[y,x] = 0
+                else:
+                    self.matrix_br_corners_rows[y][x] = index_br[0]
+                    self.matrix_br_corners_cols[y][x] = index_br[1]
+                    self.matrix_br_multiply[y,x] = (1.0-x_ratio)*(1.0-y_ratio)
+
+    def distort_image(self, image):
+        # index pixels to get the 4 corners for bilinear combination
+        tl_pixels = image[self.matrix_tl_corners_rows, self.matrix_tl_corners_cols]
+        tr_pixels = image[self.matrix_tr_corners_rows, self.matrix_tr_corners_cols]
+        bl_pixels = image[self.matrix_bl_corners_rows, self.matrix_bl_corners_cols]
+        br_pixels = image[self.matrix_br_corners_rows, self.matrix_br_corners_cols]
+
+        # bilinear ratios, elemwise multiply
+        tl_pixels = numpy.multiply(tl_pixels, self.matrix_tl_multiply)
+        tr_pixels = numpy.multiply(tr_pixels, self.matrix_tr_multiply)
+        bl_pixels = numpy.multiply(bl_pixels, self.matrix_bl_multiply)
+        br_pixels = numpy.multiply(br_pixels, self.matrix_br_multiply)
+
+        # sum to finish bilinear combination
+        return numpy.sum([tl_pixels,tr_pixels,bl_pixels,br_pixels], axis=0)
+
+# TESTS ----------------------------------------------------------------------
+
+def _load_image(filepath):
+    _RGB_TO_GRAYSCALE = [0.3, 0.59, 0.11, 0.0]
+    img = Image.open(filepath)
+    img = numpy.asarray(img)
+    if len(img.shape) > 2:
+        img = (img * _RGB_TO_GRAYSCALE).sum(axis=2)
+    return (img / 255.0).astype('float')
+
+def _specific_test():
+    img = _load_image("tests/d.png")
+    dist = LocalElasticDistorter((32,32), (15,15), 9.0, 5.0)
+    dist.distort_image(img)
+
+def _distorter_tests():
+    #import pylab
+    #pylab.imshow(img)
+    #pylab.show()
+
+    for letter in ("d", "a", "n", "o"):
+        img = _load_image("tests/" + letter + ".png")
+        for alpha in (1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0):
+            for sigma in (1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0):
+                id = LocalElasticDistorter((32,32), (15,15), sigma, alpha)
+                img2 = id.distort_image(img)
+                img2 = Image.fromarray((img2 * 255).astype('uint8'), "L")
+                img2.save("tests/"+letter+"_alpha"+str(alpha)+"_sigma"+str(sigma)+".png")
+
+def _benchmark():
+    img = _load_image("tests/d.png")
+    dist = LocalElasticDistorter((32,32), (10,10), 5.0, 5.0)
+    import time
+    t1 = time.time()
+    for i in range(10000):
+        if i % 1000 == 0:
+            print "-"
+        dist.distort_image(img)
+    t2 = time.time()
+    print "t2-t1", t2-t1
+    print "avg", 10000/(t2-t1)
+
+if __name__ == '__main__':
+    import Image
+    _distorter_tests()
+    #_benchmark()
+    #_specific_test()
+
+