[extractor] added FFT features, modified Freq Features (Li), added LTSS features

bob/pad/face/extractor/FFTFeatures.py

@@ -26,12 +26,13 @@ class FFTFeatures(Extractor, object):
   debug: boolean
     Plot stuff
   """
-  def __init__(self, framerate=25, nfft=256, debug=False, **kwargs):
+  def __init__(self, framerate=25, nfft=256, concat=False, debug=False, **kwargs):
 
     super(FFTFeatures, self).__init__(**kwargs)
     
     self.framerate = framerate
     self.nfft = nfft
+    self.concat = concat
     self.debug = debug
 
 
@@ -72,7 +73,11 @@ class FFTFeatures(Extractor, object):
       ffts = numpy.zeros((3, output_dim))
       for i in range(3):
         ffts[i] = abs(numpy.fft.rfft(signal[:, i], n=self.nfft))
-      fft = ffts[1]
+
+      if self.concat:
+        fft = numpy.concatenate([ffts[0], ffts[1], ffts[2]])
+      else:
+        fft = ffts[1]
       
     if self.debug: 
       from matplotlib import pyplot

bob/pad/face/extractor/FreqFeatures.py

@@ -8,7 +8,6 @@ from bob.bio.base.extractor import Extractor
 import logging
 logger = logging.getLogger("bob.pad.face")
 
-from scipy.signal import welch
 
 
 class FreqFeatures(Extractor, object):

bob/pad/face/extractor/LTSS.py

+#!/usr/bin/env python
+# encoding: utf-8
+
+import numpy
+
+from bob.bio.base.extractor import Extractor
+
+import logging
+logger = logging.getLogger("bob.pad.face")
+
+from scipy.fftpack import rfft
+
+
+class LTSS(Extractor, object):
+  """
+  Compute Long-term spectral statistics of a pulse signal.
+  
+  The features are described in the following article:
+  
+    @Article {
+    Author         = {Muckenhirn, Hannah and Korshunov, Pavel and
+                     Magimai-Doss, Mathew and Marcel, Sebastien }
+    Title          = {Long-Term Spectral Statistics for Voice Presentation
+                     Attack Detection},
+    Journal        = {IEEE/ACM Trans. Audio, Speech and Lang. Proc.},
+    Volume         = {25},
+    Number         = {11},
+    Pages          = {2098--2111},
+    year           = 2017
+    }
+
+  **Parameters:**
+
+  framerate: int
+    The sampling frequency of the signal (i.e the framerate ...) 
+
+  nfft: int
+    Number of points to compute the FFT
+
+  debug: boolean
+    Plot stuff
+  """
+  def __init__(self, window_size=25, framerate=25, nfft=64, concat=False, debug=False, **kwargs):
+
+    super(LTSS, self).__init__()
+    
+    self.framerate = framerate
+    self.nfft = nfft
+    self.debug = debug
+    self.window_size = window_size
+    self.concat = concat
+
+  def _get_ltss(self, signal):
+    
+    # log-magnitude of DFT coefficients
+    log_mags = []
+    window_stride = int(self.window_size / 2)
+   
+    # go through windows
+    for w in range(0, (signal.shape[0] - self.window_size), window_stride):
+      fft = rfft(signal[w:w+self.window_size], n=self.nfft)
+      mags = numpy.zeros(int(self.nfft/2), dtype=numpy.float64)
+      mags[0] = abs(fft[0])
+      index = 1
+      for i in range(1, (fft.shape[0]-1), 2):
+        mags[index] = numpy.sqrt(fft[i]**2 + fft[i+1]**2)
+        if mags[index] < 1:
+          mags[index] = 1
+        index += 1
+      log_mags.append(numpy.log(mags))
+
+    # get the long term statistics
+    log_mags = numpy.array(log_mags)
+    mean = numpy.mean(log_mags, axis=0)
+    std = numpy.std(log_mags, axis=0)
+    ltss = numpy.concatenate([mean, std])
+    return ltss
+
+
+  def __call__(self, signal):
+    """
+    Computes the long-term spectral statistics for a given signal.
+
+    **Parameters**
+
+    signal: numpy.array 
+      The signal
+
+    **Returns:**
+
+      feature: numpy.array 
+       the long-term spectral statistics feature vector 
+    """
+    # sanity check
+    if signal.ndim == 1:
+      if numpy.isnan(numpy.sum(signal)):
+        return
+    if signal.ndim == 2 and (signal.shape[1] == 3):
+      for i in range(signal.shape[1]):
+        if numpy.isnan(numpy.sum(signal[:, i])):
+          return
+
+    if signal.ndim == 1:
+      feature = self._get_ltss(signal)
+
+    if signal.ndim == 2 and (signal.shape[1] == 3):
+      
+      if not self.concat:
+        feature = self._get_ltss(signal[:, 1])
+      else:
+        ltss = []
+        for i in range(signal.shape[1]):
+          ltss.append(self._get_ltss(signal[:, i]))
+        feature = numpy.concatenate([ltss[0], ltss[1], ltss[2]])
+
+    if numpy.isnan(numpy.sum(feature)):
+      logger.warn("Feature not extracted")
+      return
+    if numpy.sum(feature) == 0:
+      logger.warn("Feature not extracted")
+      return
+
+    return feature

bob/pad/face/extractor/__init__.py

@@ -9,6 +9,8 @@ from .FrequencySpectrum import FrequencySpectrum
 from .FreqFeatures import FreqFeatures
 from .NormalizeLength import NormalizeLength
 from .FFTFeatures import FFTFeatures 
+from .LTSS import LTSS 
+from .Autocorrelation import Autocorrelation 
 
 def __appropriate__(*args):
     """Says object was actually declared here, and not in the import module.