最终发现了问题——我使用 NormalBayesClassifier 错误。它并不意味着直接输入数十张高清图像：首先应该使用 OpenCV 的其他算法来处理它们。
我最终做了以下事情： + 将图像裁剪到可能包含对象的区域 + 将图像转换为灰度 + 使用 cv2.goodFeaturesToTrack() 从裁剪区域收集特征以训练分类器。
少数特征对我有用，也许是因为我已经裁剪了图像，并且它很幸运地包含了在一个类中被遮挡的高对比度对象。
以下代码获得了高达 95% 的人口正确率：
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import cv2
import sys, os.path, getopt
import numpy, random
def _usage():
print
print "cvbayes trainer"
print
print "Options:"
print
print "-m --ham= path to dir of ham images"
print "-s --spam= path to dir of spam images"
print "-h --help this help text"
print "-v --verbose lots more output"
print
def _parseOpts(argv):
"""
Turn options + args into a dict of config we'll follow. Merge in default conf.
"""
try:
opts, args = getopt.getopt(argv[1:], "hm:s:v", ["help", "ham=", 'spam=', 'verbose'])
except getopt.GetoptError as err:
print(err) # will print something like "option -a not recognized"
_usage()
sys.exit(2)
optsDict = {}
for o, a in opts:
if o == "-v":
optsDict['verbose'] = True
elif o in ("-h", "--help"):
_usage()
sys.exit()
elif o in ("-m", "--ham"):
optsDict['ham'] = a
elif o in ('-s', '--spam'):
optsDict['spam'] = a
else:
assert False, "unhandled option"
for mandatory_arg in ('ham', 'spam'):
if mandatory_arg not in optsDict:
print "Mandatory argument '%s' was missing; cannot continue" % mandatory_arg
sys.exit(0)
return optsDict
class ClassifierWrapper(object):
"""
Setup and encapsulate a naive bayes classifier based on OpenCV's
NormalBayesClassifier. Presently we do not use it intelligently,
instead feeding in flattened arrays of B&W pixels.
"""
def __init__(self):
super(ClassifierWrapper,self).__init__()
self.classifier = cv2.NormalBayesClassifier()
self.data = []
self.responses = []
def _load_image_features(self, f):
image_colour = cv2.imread(f)
image_crop = image_colour[327:390, 784:926] # Use the junction boxes, luke
image_grey = cv2.cvtColor(image_crop, cv2.COLOR_BGR2GRAY)
features = cv2.goodFeaturesToTrack(image_grey, 4, 0.02, 3)
return features.flatten()
def train_from_file(self, f, cl):
features = self._load_image_features(f)
self.data.append(features)
self.responses.append(cl)
def train(self, update=False):
matrix_data = numpy.matrix( self.data ).astype('float32')
matrix_resp = numpy.matrix( self.responses ).astype('float32')
self.classifier.train(matrix_data, matrix_resp, update=update)
self.data = []
self.responses = []
def predict_from_file(self, f):
features = self._load_image_features(f)
features_matrix = numpy.matrix( [ features ] ).astype('float32')
retval, results = self.classifier.predict( features_matrix )
return results
if __name__ == "__main__":
opts = _parseOpts(sys.argv)
cw = ClassifierWrapper()
ham = os.listdir(opts['ham'])
spam = os.listdir(opts['spam'])
n_training_samples = min( [len(ham),len(spam)])
print "Will train on %d samples for equal sets" % n_training_samples
for f in random.sample(ham, n_training_samples):
img_path = os.path.join(opts['ham'], f)
print "ham: %s" % img_path
cw.train_from_file(img_path, 2)
for f in random.sample(spam, n_training_samples):
img_path = os.path.join(opts['spam'], f)
print "spam: %s" % img_path
cw.train_from_file(img_path, 1)
cw.train()
print
print
# spam dir much bigger so mostly unused, let's try predict() on all of it
print "predicting on all spam..."
n_wrong = 0
n_files = len(os.listdir(opts['spam']))
for f in os.listdir(opts['spam']):
img_path = os.path.join(opts['spam'], f)
result = cw.predict_from_file(img_path)
print "%s\t%s" % (result, img_path)
if result[0][0] == 2:
n_wrong += 1
print
print "got %d of %d wrong = %.1f%%" % (n_wrong, n_files, float(n_wrong)/n_files * 100, )
现在我正在用垃圾邮件的一个随机子集对其进行训练，仅仅是因为它的数量要多得多，而且每个类的训练数据量应该大致相等。如果数据更好（例如，当光照不同时，总是包括黎明和黄昏的样本），它可能会更高。
也许即使是 NormalBayesClassifier 也不是这项工作的错误工具，我应该尝试跨连续帧进行运动检测 - 但至少互联网现在有一个示例可供挑选。