def clusteringParamSearch(trainPath, trainClasses, trainCount, clusterRange, ptRange, etRange, fisherRange):
trainImgs = im.loadImages(trainPath, trainClasses, trainCount)
trainLabels = im.loadLabels(trainPath, trainClasses, trainCount)
for peakVal in ptRange:
for edgeVal in etRange:
# Extract SIFT features
rec.logger.info("SIFT Parameter. Peak: %f Edge: %f", peakVal, edgeVal)
trainDesc, numTrainingDesc = im.extractFeatures(trainImgs, peakVal, edgeVal)
for clusterParam in clusterRange:
for fDist in fisherRange:
learner = cl.Learner()
learner.peakThreshold = peakVal
learner.edgeThreshold = edgeVal
learner.numTrainingDesc = numTrainingDesc
rec.logger.info("Cluster parameter. Cluster: %i MinDist %f", clusterParam, fDist)
learner.averageSilhouetteCoefficient(trainLabels, trainDesc, clusterParam, fDist)
learner.saveToDB()
def train(self, featOrImgs, labels, numCluster, peakThreshold=10.0, edgeThreshold=10.0, maxiter=0, numruns=20):
"""
Perform k-means clustering and vector quantize all features.
The codebook and the histograms of visual words are stored as members.
If L{featOrImgs} is a list of images, SIFT features are extracted
before all other operations are done.
@param featOrImgs: Either a list of SIFT descriptor arrays or a list of images.
@type featOrImgs: [numpy.ndarray] or [Image.Image]
@param labels: A list of labels corresponding to the list of descriptors/images.
@type labels: [string]
@param numCluster: The number of clusters used for k-means clustering.
@type numCluster: int
@param peakThreshold: The peak threshold used during SIFT feature extraction.
This argument is ignored if L{featOrImgs} is a list of SIFT features.
@type peakThreshold: float
@param edgeThreshold: The edge threshold used during SIFT feature extraction.
This argument is ignored if L{featOrImgs} is a list of SIFT features.
@type edgeThreshold: float
@param maxiter: The maximum iterations of one k-means run.
"0" means there is no iteration limit.
@type maxiter: int
@param numruns: The number of k-means runs.
@type numruns: int
"""
self.numCluster = numCluster
if isinstance(featOrImgs[0], Image.Image):
self.peakThreshold = peakThreshold
self.edgeThreshold = edgeThreshold
descriptors = im.extractFeatures(featOrImgs, peakThreshold, edgeThreshold)
else:
descriptors = featOrImgs
starttime = time.time()
# Compute codebook from all sift descriptors of all images
self.codebook = self._cluster(descriptors, numCluster, maxiter, numruns)
# Get the histograms as a list of numpy arrays
tempHistograms = _buildHistograms(descriptors)
# Create orange domain
classes = _getAllClasses(labels)
self.strClasses = str(classes)
histogramDomain = [orange.FloatVariable('a%i' % x) for x in xrange(len(tempHistograms[0]))]
classDomain = [orange.EnumVariable("class", values=orange.StringList(classes))]
self.domain = orange.Domain(histogramDomain + classDomain)
# Create orange exampleTable
self.histograms = _convertToOrangeDataSet(tempHistograms, self.domain, labels)
endtime = time.time()
self.trainingTime = endtime - starttime
def classificationAccuracy(self, histOrImgs, labels=None, confThr=0.0,
peakThr=None, edgeThr=None, nu=0.6, gamma=2.0, doCrossVal=False):
"""
Classify test data and (optionally) perform a cross validation on the training data.
@param histOrImgs: Either a list of SIFT descriptor arrays or an iterator of images.
@type histOrImgs: numpy.ndarray or [numpy.ndarray] or Image.Image or [Image.Image]
@param labels: A list of labels corresponding to the list of descriptors/images.
@type labels: [string]
@param nu: The S{nu}-Parameter of the support vector machine.
@type nu: float
@param gamma: The S{gamma}-Parameter of the RBF-Kernel.
@type gamma: float
@param confidenceThreshold: All classifications who are classified with a lower
confidence than this threshold are rejected. 1.0:
Everything is rejected, 0.0: Nothing is rejected.
@type confidenceThreshold: float
@param peakThreshold: A SIFT parameter. Sensible values: 0.0 < x < 30.0.
@type peakThreshold: float
@param edgeThreshold: A SIFT parameter. Sensible values: 0.0 < x < 10.0.
@type edgeThreshold: float
@rtype: (float,float)
@return: The cross validation accuracy and the test data classification accuracy.
"""
if self.learner is None:
raise ValueError("Learner has to be loaded before classification can be done.")
# Set SIFT member variables (so they get stored in the DB if requested)
if peakThr is None:
self.peakThreshold = self.learner.peakThreshold
else:
self.peakThreshold = peakThr
if edgeThr is None:
self.edgeThreshold = self.learner.edgeThreshold
else:
self.edgeThreshold = edgeThr
# If we've been given an images iterator, extract features and vector quantize
if isinstance(histOrImgs, collections.Iterator):
if labels is None:
raise ValueError("If argument 'histOrImgs' is an iterator of images, \
the argument 'lables' must not be None.")
desc, self.numTestDesc = im.extractFeatures(histOrImgs, self.peakThreshold, self.edgeThreshold)
recognosco.logger.info("Found %i features/image on average.", self.numTestDesc / len(desc))
tmpHistograms = _buildHistograms(self.learner.codebook, desc)
histograms = _convertToOrangeDataSet(tmpHistograms, self.learner.domain, labels)
else:
histograms = histOrImgs
values = histograms.domain.classVar.values
self.values = values
length = len(values)
self.confusion = numpy.zeros((length, length), int)
starttime = time.time()
self.nu = nu
self.gamma = gamma
svm = orange.SVMLearner()
svm.svm_type = orange.SVMLearner.Nu_SVC
svm.nu = nu
svm.gamma = gamma
svm.kernel_type = orange.SVMLearner.RBF
svm.probability = True
recognosco.logger.debug("Training Support Vector Machine...")
self.classifier = svm(self.learner.histograms)
recognosco.logger.debug("Done...")
crossVal = 0.0
if doCrossVal:
crossVal = orngTest.crossValidation([svm], self.learner.histograms, folds=10)
numCorrectClassified = 0.0
numClassified = 0.0
for i in range(len(histograms)):
c = self.classifier(histograms[i])
recognosco.logger.info("Has the Class: %s", histograms[i].getclass())
recognosco.logger.info("Classified as: %s", c)
prob = self.classifier(histograms[i], self.classifier.GetProbabilities)
conf = self.__getConfidence(prob)
recognosco.logger.info("Confidence: %f", conf)
if conf < confThr:
recognosco.logger.info("Rejected classification (Threshold: %.2f)", confThr)
continue
numClassified += 1.0
predicted = values.index(str(c))
actual = values.index(str(histograms[i].getclass()))
self.confusion[predicted][actual] += 1
if(c == histograms[i].getclass()):
numCorrectClassified += 1.0
endtime = time.time()
self.confusion = str(self.confusion)
self.clAccuracy = numCorrectClassified / numClassified
self.fracClassified = numClassified / len(histograms)
if doCrossVal:
self.cvAccuracy = orngStat.CA(crossVal)[0]
recognosco.logger.info("Cross validation accuracy: %s", self.cvAccuracy)
else:
self.cvAccuracy = -1.0
recognosco.logger.info("Classification accuracy of test data: %s", self.clAccuracy)
self.testTime = endtime - starttime
return (self.cvAccuracy, self.classificationAccuracy)
def classify(self, images, nu=0.6, gamma=4.0, confidenceThreshold=0.03, minFeatures=5):
"""
Classify test data and return the label(s).
@param images: A single image or a list of images.
@type images: Image.Image or [Image.Image]
@param nu: The S{nu}-Parameter of the support vector machine.
@type nu: float
@param gamma: The S{gamma}-Parameter of the RBF-Kernel..
@type gamma: float
@param confidenceThreshold: All classifications who are classified with a lower
confidence than this threshold are rejected. 1.0:
Everything is rejected, 0.0: Nothing is rejected.
@type confidenceThreshold: float
@param gamma: The minimum amount of features which should be extracted from
every image.
@type gamma: float
@return: The labels which result from the classification
@rtype: [string]
"""
if self.learner is None:
raise ValueError("Learner has to be loaded before classification can be done.")
if not isinstance(images, collections.Iterable):
images = [images]
# We have no labels, so set them to None
inputLabels = [None for i in range(len(images))]
# Use same SIFT parameters as during training, but change SIFT parameters
# if they result in too less features
self.numTestDesc = 0
peak = self.learner.peakThreshold
edge = self.learner.edgeThreshold
peakDecr = 1.0
edgeInc = 1.0
descIter1, descIter2 = itertools.tee(im.extractFeatures(images, peak, edge))
for desc in descIter1:
self.numTestDesc += numpy.size(desc, 0)
recognosco.logger.info("Found %i features.", self.numTestDesc)
while self.numTestDesc < minFeatures and peak >= 0.0 and edge <= 10.0:
# Adjust SIFT parameters
peakDecr = peakDecr - 0.2
peak = self.learner.peakThreshold * peakDecr
edgeInc = edgeInc + 0.2
edge = self.learner.peakThreshold * edgeInc
recognosco.logger.warning("Less than %i SIFT features have been extracted. "
"Adjusting SIFT parameters (peak: %f, edge: %f).",
minFeatures, peak, edge)
descIter1, descIter2 = itertools.tee(im.extractFeatures(images, peak, edge))
self.numTestDesc = 0
for desc in descIter1:
self.numTestDesc += numpy.size(desc, 0)
recognosco.logger.info("Found %i features.", self.numTestDesc)
if self.numTestDesc < minFeatures:
recognosco.logger.error("Less than %i SIFT features have been extracted. Aborting.", minFeatures)
return ["<not enough features>"]
tmpHistograms = _buildHistograms(self.learner.codebook, descIter2)
histograms = _convertToOrangeDataSet(tmpHistograms, self.learner.domain, inputLabels)
self.nu = nu
self.gamma = gamma
svm = orange.SVMLearner()
svm.svm_type = orange.SVMLearner.Nu_SVC
svm.nu = nu
svm.gamma = gamma
svm.kernel_type = orange.SVMLearner.RBF
svm.probability = True
classifier = svm(self.learner.histograms)
outputLabels = []
for i in range(len(histograms)):
c = classifier(histograms[i])
prob = classifier(histograms[i], classifier.GetProbabilities)
conf = self.__getConfidence(prob)
if conf < confidenceThreshold:
recognosco.logger.info("Rejected classification. Nearest match:"
" %s (Confidence: %.2f%%)", c, conf * 100.0)
continue
recognosco.logger.info("Classified as: %s. Confidence: %.2f%%",
c, conf * 100.0)
outputLabels.append(str(c))
return outputLabels
