test.py 20.3 KB
Newer Older
Pedro TOME's avatar
Pedro TOME committed
1 2 3 4
#!/usr/bin/env python
# vim: set fileencoding=utf-8 :


5 6 7 8 9 10 11 12
"""Unit tests against references extracted from

Matlab code from Bram Ton available on the matlab central website:

https://www.mathworks.com/matlabcentral/fileexchange/35754-wide-line-detector

This code implements the detector described in [HDLTL10] (see the references in
the generated sphinx documentation)
Pedro TOME's avatar
Pedro TOME committed
13 14 15
"""

import os
16
import numpy
17
import numpy as np
18 19
import nose.tools

Pedro TOME's avatar
Pedro TOME committed
20 21
import pkg_resources

22 23
import bob.io.base
import bob.io.matlab
André Anjos's avatar
André Anjos committed
24
import bob.io.image
Pedro TOME's avatar
Pedro TOME committed
25

26
from ..preprocessor import utils as preprocessor_utils
27

Pedro TOME's avatar
Pedro TOME committed
28

29 30 31 32
def F(parts):
  """Returns the test file path"""

  return pkg_resources.resource_filename(__name__, os.path.join(*parts))
André Anjos's avatar
André Anjos committed
33

34 35 36

def test_finger_crop():

André Anjos's avatar
André Anjos committed
37 38
  input_filename = F(('preprocessors', '0019_3_1_120509-160517.png'))
  output_img_filename  = F(('preprocessors',
39
    '0019_3_1_120509-160517_img_lee_huang.mat'))
André Anjos's avatar
André Anjos committed
40
  output_fvr_filename  = F(('preprocessors',
41 42 43 44
    '0019_3_1_120509-160517_fvr_lee_huang.mat'))

  img = bob.io.base.load(input_filename)

45 46 47 48 49 50 51 52 53
  from bob.bio.vein.preprocessor import Preprocessor, LeeMask, \
      HuangNormalization, NoFilter

  processor = Preprocessor(
      LeeMask(filter_height=40, filter_width=4),
      HuangNormalization(padding_width=0, padding_constant=0),
      NoFilter(),
      )
  preproc, mask = processor(img)
54
  #preprocessor_utils.show_mask_over_image(preproc, mask)
55

56 57 58 59 60
  mask_ref = bob.io.base.load(output_fvr_filename).astype('bool')
  preproc_ref = bob.core.convert(bob.io.base.load(output_img_filename),
      numpy.uint8, (0,255), (0.0,1.0))

  assert numpy.mean(numpy.abs(mask - mask_ref)) < 1e-2
61

62
  # Very loose comparison!
63
  #preprocessor_utils.show_image(numpy.abs(preproc.astype('int16') - preproc_ref.astype('int16')).astype('uint8'))
64
  assert numpy.mean(numpy.abs(preproc - preproc_ref)) < 1.3e2
65 66


67
def test_max_curvature():
68 69 70

  #Maximum Curvature method against Matlab reference

71 72 73 74 75 76 77 78 79 80 81 82
  image = bob.io.base.load(F(('extractors', 'image.hdf5')))
  image = image.T
  image = image.astype('float64')/255.
  mask  = bob.io.base.load(F(('extractors', 'mask.hdf5')))
  mask  = mask.T
  mask  = mask.astype('bool')
  vt_ref = bob.io.base.load(F(('extractors', 'mc_vt_matlab.hdf5')))
  vt_ref = vt_ref.T
  g_ref = bob.io.base.load(F(('extractors', 'mc_g_matlab.hdf5')))
  g_ref = g_ref.T
  bin_ref = bob.io.base.load(F(('extractors', 'mc_bin_matlab.hdf5')))
  bin_ref = bin_ref.T
83 84

  # Apply Python implementation
85
  from bob.bio.vein.extractor.MaximumCurvature import MaximumCurvature
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105
  MC = MaximumCurvature(3) #value used to create references

  kappa = MC.detect_valleys(image, mask)
  V = MC.eval_vein_probabilities(kappa)
  Vt = V.sum(axis=2)
  Cd = MC.connect_centres(Vt)
  G = numpy.amax(Cd, axis=2)
  bina = MC.binarise(G)

  assert numpy.allclose(Vt, vt_ref, 1e-3, 1e-4), \
      'Vt differs from reference by %s' % numpy.abs(Vt-vt_ref).sum()
  # Note: due to Matlab implementation bug, can only compare in a limited
  # range with a 3-pixel around frame
  assert numpy.allclose(G[2:-3,2:-3], g_ref[2:-3,2:-3]), \
      'G differs from reference by %s' % numpy.abs(G-g_ref).sum()
  # We require no more than 30 pixels (from a total of 63'840) are different
  # between ours and the matlab implementation
  assert numpy.abs(bin_ref-bina).sum() < 30, \
      'Binarized image differs from reference by %s' % \
      numpy.abs(bin_ref-bina).sum()
Pedro TOME's avatar
Pedro TOME committed
106 107


Vedrana KRIVOKUCA's avatar
Vedrana KRIVOKUCA committed
108 109 110 111 112 113
def test_max_curvature_HE():
  # Maximum Curvature method when Histogram Equalization post-processing is applied to the preprocessed vein image

  # Read in input image
  input_img_filename = F(('preprocessors', '0019_3_1_120509-160517.png'))
  input_img = bob.io.base.load(input_img_filename)
André Anjos's avatar
André Anjos committed
114

Vedrana KRIVOKUCA's avatar
Vedrana KRIVOKUCA committed
115
  # Preprocess the data and apply Histogram Equalization postprocessing (same parameters as in maximum_curvature.py configuration file + postprocessing)
116 117 118 119 120 121 122 123
  from bob.bio.vein.preprocessor import Preprocessor, LeeMask, \
      HuangNormalization, HistogramEqualization
  processor = Preprocessor(
      LeeMask(filter_height=40, filter_width=4),
      HuangNormalization(padding_width=0, padding_constant=0),
      HistogramEqualization(),
      )
  preproc_data = processor(input_img)
Vedrana KRIVOKUCA's avatar
Vedrana KRIVOKUCA committed
124 125 126 127 128 129 130

  # Extract features from preprocessed and histogram equalized data using MC extractor (same parameters as in maximum_curvature.py configuration file)
  from bob.bio.vein.extractor.MaximumCurvature import MaximumCurvature
  MC = MaximumCurvature(sigma = 5)
  extr_data = MC(preproc_data)


131
def test_repeated_line_tracking():
132 133 134

  #Repeated Line Tracking method against Matlab reference

André Anjos's avatar
André Anjos committed
135 136 137
  input_img_filename  = F(('extractors', 'miurarlt_input_img.mat'))
  input_fvr_filename  = F(('extractors', 'miurarlt_input_fvr.mat'))
  output_filename     = F(('extractors', 'miurarlt_output.mat'))
138 139 140 141 142 143

  # Load inputs
  input_img = bob.io.base.load(input_img_filename)
  input_fvr = bob.io.base.load(input_fvr_filename)

  # Apply Python implementation
144
  from bob.bio.vein.extractor.RepeatedLineTracking import RepeatedLineTracking
145 146 147 148 149 150 151 152 153 154 155
  RLT = RepeatedLineTracking(3000, 1, 21, False)
  output_img = RLT((input_img, input_fvr))

  # Load Matlab reference
  output_img_ref = bob.io.base.load(output_filename)

  # Compare output of python's implementation to matlab reference
  # (loose comparison!)
  assert numpy.mean(numpy.abs(output_img - output_img_ref)) < 0.5


Vedrana KRIVOKUCA's avatar
Vedrana KRIVOKUCA committed
156 157 158 159 160 161
def test_repeated_line_tracking_HE():
  # Repeated Line Tracking method when Histogram Equalization post-processing is applied to the preprocessed vein image

  # Read in input image
  input_img_filename = F(('preprocessors', '0019_3_1_120509-160517.png'))
  input_img = bob.io.base.load(input_img_filename)
André Anjos's avatar
André Anjos committed
162

Vedrana KRIVOKUCA's avatar
Vedrana KRIVOKUCA committed
163
  # Preprocess the data and apply Histogram Equalization postprocessing (same parameters as in repeated_line_tracking.py configuration file + postprocessing)
164 165 166 167 168 169 170 171
  from bob.bio.vein.preprocessor import Preprocessor, LeeMask, \
      HuangNormalization, HistogramEqualization
  processor = Preprocessor(
      LeeMask(filter_height=40, filter_width=4),
      HuangNormalization(padding_width=0, padding_constant=0),
      HistogramEqualization(),
      )
  preproc_data = processor(input_img)
Vedrana KRIVOKUCA's avatar
Vedrana KRIVOKUCA committed
172 173 174 175 176 177 178 179 180 181 182 183 184

  # Extract features from preprocessed and histogram equalized data using RLT extractor (same parameters as in repeated_line_tracking.py configuration file)
  from bob.bio.vein.extractor.RepeatedLineTracking import RepeatedLineTracking
  # Maximum number of iterations
  NUMBER_ITERATIONS = 3000
  # Distance between tracking point and cross section of profile
  DISTANCE_R = 1
  # Width of profile
  PROFILE_WIDTH = 21
  RLT = RepeatedLineTracking(iterations = NUMBER_ITERATIONS, r = DISTANCE_R, profile_w = PROFILE_WIDTH, seed = 0)
  extr_data = RLT(preproc_data)


185
def test_wide_line_detector():
186 187 188

  #Wide Line Detector method against Matlab reference

André Anjos's avatar
André Anjos committed
189 190 191
  input_img_filename  = F(('extractors', 'huangwl_input_img.mat'))
  input_fvr_filename  = F(('extractors', 'huangwl_input_fvr.mat'))
  output_filename     = F(('extractors', 'huangwl_output.mat'))
192 193 194 195 196 197

  # Load inputs
  input_img = bob.io.base.load(input_img_filename)
  input_fvr = bob.io.base.load(input_fvr_filename)

  # Apply Python implementation
198
  from bob.bio.vein.extractor.WideLineDetector import WideLineDetector
199 200 201 202 203 204 205 206 207 208
  WL = WideLineDetector(5, 1, 41, False)
  output_img = WL((input_img, input_fvr))

  # Load Matlab reference
  output_img_ref = bob.io.base.load(output_filename)

  # Compare output of python's implementation to matlab reference
  assert numpy.allclose(output_img, output_img_ref)


Vedrana KRIVOKUCA's avatar
Vedrana KRIVOKUCA committed
209 210 211 212 213 214
def test_wide_line_detector_HE():
  # Wide Line Detector method when Histogram Equalization post-processing is applied to the preprocessed vein image

  # Read in input image
  input_img_filename = F(('preprocessors', '0019_3_1_120509-160517.png'))
  input_img = bob.io.base.load(input_img_filename)
André Anjos's avatar
André Anjos committed
215

Vedrana KRIVOKUCA's avatar
Vedrana KRIVOKUCA committed
216
  # Preprocess the data and apply Histogram Equalization postprocessing (same parameters as in wide_line_detector.py configuration file + postprocessing)
217 218 219 220 221 222 223 224
  from bob.bio.vein.preprocessor import Preprocessor, LeeMask, \
      HuangNormalization, HistogramEqualization
  processor = Preprocessor(
      LeeMask(filter_height=40, filter_width=4),
      HuangNormalization(padding_width=0, padding_constant=0),
      HistogramEqualization(),
      )
  preproc_data = processor(input_img)
Vedrana KRIVOKUCA's avatar
Vedrana KRIVOKUCA committed
225 226 227 228 229 230 231 232 233 234 235 236 237

  # Extract features from preprocessed and histogram equalized data using WLD extractor (same parameters as in wide_line_detector.py configuration file)
  from bob.bio.vein.extractor.WideLineDetector import WideLineDetector
  # Radius of the circular neighbourhood region
  RADIUS_NEIGHBOURHOOD_REGION = 5
  NEIGHBOURHOOD_THRESHOLD = 1
  # Sum of neigbourhood threshold
  SUM_NEIGHBOURHOOD = 41
  RESCALE = True
  WLD = WideLineDetector(radius = RADIUS_NEIGHBOURHOOD_REGION, threshold = NEIGHBOURHOOD_THRESHOLD, g = SUM_NEIGHBOURHOOD, rescale = RESCALE)
  extr_data = WLD(preproc_data)


238 239
def test_miura_match():

André Anjos's avatar
André Anjos committed
240 241 242 243 244
  #Match Ratio method against Matlab reference

  template_filename = F(('algorithms', '0001_2_1_120509-135338.mat'))
  probe_gen_filename = F(('algorithms', '0001_2_2_120509-135558.mat'))
  probe_imp_filename = F(('algorithms', '0003_2_1_120509-141255.mat'))
245 246 247 248 249

  template_vein = bob.io.base.load(template_filename)
  probe_gen_vein = bob.io.base.load(probe_gen_filename)
  probe_imp_vein = bob.io.base.load(probe_imp_filename)

250
  from bob.bio.vein.algorithm.MiuraMatch import MiuraMatch
251 252 253 254 255 256 257
  MM = MiuraMatch(ch=18, cw=28)
  score_gen = MM.score(template_vein, probe_gen_vein)

  assert numpy.isclose(score_gen, 0.382689335394127)

  score_imp = MM.score(template_vein, probe_imp_vein)
  assert numpy.isclose(score_imp, 0.172906739278421)
258 259 260 261 262 263 264


def test_assert_points():

  # Tests that point assertion works as expected
  area = (10, 5)
  inside = [(0,0), (3,2), (9, 4)]
265
  preprocessor_utils.assert_points(area, inside) #should not raise
266 267 268 269

  def _check_outside(point):
    # should raise, otherwise it is an error
    try:
270
      preprocessor_utils.assert_points(area, [point])
271 272 273 274 275 276 277 278 279 280 281 282 283 284
    except AssertionError as e:
      assert str(point) in str(e)
    else:
      raise AssertionError("Did not assert %s is outside of %s" % (point, area))

  outside = [(-1, 0), (10, 0), (0, 5), (10, 5), (15,12)]
  for k in outside: _check_outside(k)


def test_fix_points():

  # Tests that point clipping works as expected
  area = (10, 5)
  inside = [(0,0), (3,2), (9, 4)]
285
  fixed = preprocessor_utils.fix_points(area, inside)
286 287
  assert numpy.array_equal(inside, fixed), '%r != %r' % (inside, fixed)

288
  fixed = preprocessor_utils.fix_points(area, [(-1, 0)])
289 290
  assert numpy.array_equal(fixed, [(0, 0)])

291
  fixed = preprocessor_utils.fix_points(area, [(10, 0)])
292 293
  assert numpy.array_equal(fixed, [(9, 0)])

294
  fixed = preprocessor_utils.fix_points(area, [(0, 5)])
295 296
  assert numpy.array_equal(fixed, [(0, 4)])

297
  fixed = preprocessor_utils.fix_points(area, [(10, 5)])
298 299
  assert numpy.array_equal(fixed, [(9, 4)])

300
  fixed = preprocessor_utils.fix_points(area, [(15, 12)])
301 302 303 304 305 306 307 308
  assert numpy.array_equal(fixed, [(9, 4)])


def test_poly_to_mask():

  # Tests we can generate a mask out of a polygon correctly
  area = (10, 9) #10 rows, 9 columns
  polygon = [(2, 2), (2, 7), (7, 7), (7, 2)] #square shape, (y, x) format
309
  mask = preprocessor_utils.poly_to_mask(area, polygon)
310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327
  nose.tools.eq_(mask.dtype, numpy.bool)

  # This should be the output:
  expected = numpy.array([
      [False, False, False, False, False, False, False, False, False],
      [False, False, False, False, False, False, False, False, False],
      [False, False, True,  True,  True,  True,  True,  True,  False],
      [False, False, True,  True,  True,  True,  True,  True,  False],
      [False, False, True,  True,  True,  True,  True,  True,  False],
      [False, False, True,  True,  True,  True,  True,  True,  False],
      [False, False, True,  True,  True,  True,  True,  True,  False],
      [False, False, True,  True,  True,  True,  True,  True,  False],
      [False, False, False, False, False, False, False, False, False],
      [False, False, False, False, False, False, False, False, False],
      ])
  assert numpy.array_equal(mask, expected)

  polygon = [(3, 2), (5, 7), (8, 7), (7, 3)] #trapezoid, (y, x) format
328
  mask = preprocessor_utils.poly_to_mask(area, polygon)
329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354
  nose.tools.eq_(mask.dtype, numpy.bool)

  # This should be the output:
  expected = numpy.array([
      [False, False, False, False, False, False, False, False, False],
      [False, False, False, False, False, False, False, False, False],
      [False, False, False, False, False, False, False, False, False],
      [False, False, True,  False, False, False, False, False, False],
      [False, False, True,  True,  True,  False, False, False, False],
      [False, False, False, True,  True,  True,  True,  True,  False],
      [False, False, False, True,  True,  True,  True,  True,  False],
      [False, False, False, True,  True,  True,  True,  True,  False],
      [False, False, False, False, False, False, False, True,  False],
      [False, False, False, False, False, False, False, False, False],
      ])
  assert numpy.array_equal(mask, expected)


def test_mask_to_image():

  # Tests we can correctly convert a boolean array into an image
  # that makes sense according to the data types
  sample = numpy.array([False, True])
  nose.tools.eq_(sample.dtype, numpy.bool)

  def _check_uint(n):
355
    conv = preprocessor_utils.mask_to_image(sample, 'uint%d' % n)
356 357 358 359 360 361 362 363 364 365
    nose.tools.eq_(conv.dtype, getattr(numpy, 'uint%d' % n))
    target = [0, (2**n)-1]
    assert numpy.array_equal(conv, target), '%r != %r' % (conv, target)

  _check_uint(8)
  _check_uint(16)
  _check_uint(32)
  _check_uint(64)

  def _check_float(n):
366
    conv = preprocessor_utils.mask_to_image(sample, 'float%d' % n)
367 368 369 370 371 372 373 374 375 376
    nose.tools.eq_(conv.dtype, getattr(numpy, 'float%d' % n))
    assert numpy.array_equal(conv, [0, 1.0]), '%r != %r' % (conv, target)

  _check_float(32)
  _check_float(64)
  _check_float(128)


  # This should be unsupported
  try:
377
    conv = preprocessor_utils.mask_to_image(sample, 'int16')
378 379 380 381
  except TypeError as e:
    assert 'int16' in str(e)
  else:
    raise AssertionError('Conversion to int16 did not trigger a TypeError')
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396


def test_jaccard_index():

  # Tests to verify the Jaccard index calculation is accurate
  a = numpy.array([
    [False, False],
    [True, True],
    ])

  b = numpy.array([
    [True, True],
    [True, False],
    ])

397 398 399 400 401 402 403
  nose.tools.eq_(preprocessor_utils.jaccard_index(a, b), 1.0/4.0)
  nose.tools.eq_(preprocessor_utils.jaccard_index(a, a), 1.0)
  nose.tools.eq_(preprocessor_utils.jaccard_index(b, b), 1.0)
  nose.tools.eq_(preprocessor_utils.jaccard_index(a, numpy.ones(a.shape, dtype=bool)), 2.0/4.0)
  nose.tools.eq_(preprocessor_utils.jaccard_index(a, numpy.zeros(a.shape, dtype=bool)), 0.0)
  nose.tools.eq_(preprocessor_utils.jaccard_index(b, numpy.ones(b.shape, dtype=bool)), 3.0/4.0)
  nose.tools.eq_(preprocessor_utils.jaccard_index(b, numpy.zeros(b.shape, dtype=bool)), 0.0)
404 405 406 407 408 409 410 411 412 413 414 415 416 417 418


def test_intersection_ratio():

  # Tests to verify the intersection ratio calculation is accurate
  a = numpy.array([
    [False, False],
    [True, True],
    ])

  b = numpy.array([
    [True, False],
    [True, False],
    ])

419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435
  nose.tools.eq_(preprocessor_utils.intersect_ratio(a, b), 1.0/2.0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio(a, a), 1.0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio(b, b), 1.0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio(a, numpy.ones(a.shape, dtype=bool)), 1.0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio(a, numpy.zeros(a.shape, dtype=bool)), 0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio(b, numpy.ones(b.shape, dtype=bool)), 1.0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio(b, numpy.zeros(b.shape, dtype=bool)), 0)

  nose.tools.eq_(preprocessor_utils.intersect_ratio_of_complement(a, b), 1.0/2.0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio_of_complement(a, a), 0.0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio_of_complement(b, b), 0.0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio_of_complement(a, numpy.ones(a.shape, dtype=bool)), 1.0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio_of_complement(a, numpy.zeros(a.shape, dtype=bool)), 0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio_of_complement(b, numpy.ones(b.shape, dtype=bool)), 1.0)
  nose.tools.eq_(preprocessor_utils.intersect_ratio_of_complement(b, numpy.zeros(b.shape, dtype=bool)), 0)


436
def test_correlation():
437

438
  # A test for convolution performance. Correlations are used on the Miura
439 440
  # Match algorithm, therefore we want to make sure we can perform them as fast
  # as possible.
441
  import numpy
442
  import scipy.signal
443
  import bob.sp
444

445
  # Rough example from Vera fingervein database
446 447
  Y = 250
  X = 600
448 449
  CH = 80
  CW = 90
450 451 452 453 454 455

  def gen_ab():
    a = numpy.random.randint(256, size=(Y, X)).astype(float)
    b = numpy.random.randint(256, size=(Y-CH, X-CW)).astype(float)
    return a, b

456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492

  def bob_function(a, b):

    # rotate input image by 180 degrees
    b = numpy.rot90(b, k=2)

    # Determine padding size in x and y dimension
    size_a  = numpy.array(a.shape)
    size_b  = numpy.array(b.shape)
    outsize = size_a + size_b - 1

    # Determine 2D cross correlation in Fourier domain
    a2 = numpy.zeros(outsize)
    a2[0:size_a[0],0:size_a[1]] = a
    Fa = bob.sp.fft(a2.astype(numpy.complex128))

    b2 = numpy.zeros(outsize)
    b2[0:size_b[0],0:size_b[1]] = b
    Fb = bob.sp.fft(b2.astype(numpy.complex128))

    conv_ab = numpy.real(bob.sp.ifft(Fa*Fb))

    h, w = size_a - size_b + 1

    Nm = conv_ab[size_b[0]-1:size_b[0]-1+h, size_b[1]-1:size_b[1]-1+w]

    t0, s0 = numpy.unravel_index(Nm.argmax(), Nm.shape)

    # this is our output
    Nmm = Nm[t0,s0]

    # normalizes the output by the number of pixels lit on the input
    # matrices, taking into consideration the surface that produced the
    # result (i.e., the eroded model and part of the probe)
    h, w = b.shape
    return Nmm/(sum(sum(b)) + sum(sum(a[t0:t0+h-2*CH, s0:s0+w-2*CW])))

493 494

  def scipy_function(a, b):
495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510
    b = numpy.rot90(b, k=2)

    Nm = scipy.signal.convolve2d(a, b, 'valid')

    # figures out where the maximum is on the resulting matrix
    t0, s0 = numpy.unravel_index(Nm.argmax(), Nm.shape)

    # this is our output
    Nmm = Nm[t0,s0]

    # normalizes the output by the number of pixels lit on the input
    # matrices, taking into consideration the surface that produced the
    # result (i.e., the eroded model and part of the probe)
    h, w = b.shape
    return Nmm/(sum(sum(b)) + sum(sum(a[t0:t0+h-2*CH, s0:s0+w-2*CW])))

511 512

  def scipy2_function(a, b):
513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542
    b = numpy.rot90(b, k=2)
    Nm = scipy.signal.fftconvolve(a, b, 'valid')

    # figures out where the maximum is on the resulting matrix
    t0, s0 = numpy.unravel_index(Nm.argmax(), Nm.shape)

    # this is our output
    Nmm = Nm[t0,s0]

    # normalizes the output by the number of pixels lit on the input
    # matrices, taking into consideration the surface that produced the
    # result (i.e., the eroded model and part of the probe)
    h, w = b.shape
    return Nmm/(sum(sum(b)) + sum(sum(a[t0:t0+h-2*CH, s0:s0+w-2*CW])))


  def scipy3_function(a, b):
    Nm = scipy.signal.correlate2d(a, b, 'valid')

    # figures out where the maximum is on the resulting matrix
    t0, s0 = numpy.unravel_index(Nm.argmax(), Nm.shape)

    # this is our output
    Nmm = Nm[t0,s0]

    # normalizes the output by the number of pixels lit on the input
    # matrices, taking into consideration the surface that produced the
    # result (i.e., the eroded model and part of the probe)
    h, w = b.shape
    return Nmm/(sum(sum(b)) + sum(sum(a[t0:t0+h-2*CH, s0:s0+w-2*CW])))
543 544

  a, b = gen_ab()
545 546

  assert numpy.allclose(bob_function(a, b), scipy_function(a, b))
547
  assert numpy.allclose(scipy_function(a, b), scipy2_function(a, b))
548
  assert numpy.allclose(scipy2_function(a, b), scipy3_function(a, b))
549

550 551
  # if you want to test timings, uncomment the following section
  '''
552 553 554 555 556 557
  import time

  start = time.clock()
  N = 10
  for i in range(N):
    a, b = gen_ab()
558
    bob_function(a, b)
559
  total = time.clock() - start
560
  print('bob implementation, %d iterations - %.2e per iteration' % (N, total/N))
561 562 563 564 565 566

  start = time.clock()
  for i in range(N):
    a, b = gen_ab()
    scipy_function(a, b)
  total = time.clock() - start
567
  print('scipy+convolve, %d iterations - %.2e per iteration' % (N, total/N))
568 569 570 571 572 573

  start = time.clock()
  for i in range(N):
    a, b = gen_ab()
    scipy2_function(a, b)
  total = time.clock() - start
574 575 576 577 578 579 580 581 582
  print('scipy+fftconvolve, %d iterations - %.2e per iteration' % (N, total/N))

  start = time.clock()
  for i in range(N):
    a, b = gen_ab()
    scipy3_function(a, b)
  total = time.clock() - start
  print('scipy+correlate2d, %d iterations - %.2e per iteration' % (N, total/N))
  '''