def __init__(self, cfg):
BaseComponent.__init__(self, cfg)
params = cfg['params']
models_dir = params['model']
if not os.path.exists(models_dir):
raise "Error: Invalid face recognizer model directory path " + models_dir
strategies = params['strategies']
if not strategies:
raise "Error: Invalid pipeline file. Recognizer should specify atleast 1 strategy: eigen|fischer|lbp"
self.output_label = params['outputlabel']
if 'eigen' in strategies:
self.eigen = face.createEigenFaceRecognizer()
self.eigen.load(os.path.join(models_dir, 'eigen.yml'))
else:
if 'eigen' in self.output_label:
raise "Error: Invalid pipeline file. Recognizer has eigen in output label but not in strategies"
self.eigen = None
if 'fischer' in strategies:
self.fischer = face.createFisherFaceRecognizer()
self.fischer.load(os.path.join(models_dir, 'fischer.yml'))
else:
if 'fischer' in self.output_label:
raise "Error: Invalid pipeline file. Recognizer has fischer in output label but not in strategies"
self.fischer = None
if 'lbp' in strategies:
self.lbp = face.createLBPHFaceRecognizer()
self.lbp.load(os.path.join(models_dir, 'lbp.yml'))
else:
if 'lbp' in self.output_label:
raise "Error: Invalid pipeline file. Recognizer has lbp in output label but not in strategies"
self.lbp = None
with open(os.path.join(models_dir, 'model.json'), 'r') as model_file:
self.model = json.load(model_file)
self.train_img_size = (self.model['height'], self.model['width'])
self.labels = self.model['labels']
self.equalize_hist = params.get('equalizehist', False)
def recognize(img_file,
expected_label,
models_dir,
eigen=True,
fischer=True,
lbp=True,
equalize_hist=False):
eigen_label = fischer_label = lbp_label = -1
with open(os.path.join(models_dir, 'model.json'), 'r') as model_file:
model = json.load(model_file)
train_img_size = (model['height'], model['width'])
img = cv2.imread(img_file, cv2.IMREAD_GRAYSCALE)
# If training images were equalized, better to perform the same
# operation during recognition too.
if equalize_hist:
img = cv2.equalizeHist(img)
if img.shape != train_img_size:
img = cv2.resize(img, train_img_size[::-1])
if eigen:
eigen_recog = face.createEigenFaceRecognizer()
eigen_recog.load(os.path.join(models_dir, 'eigen.yml'))
eigen_label = eigen_recog.predict(img)
print('Eigen done')
if fischer:
fischer_recog = face.createFisherFaceRecognizer()
fischer_recog.load(os.path.join(models_dir, 'fischer.yml'))
fischer_label = fischer_recog.predict(img)
print('Fischer done')
if lbp:
lbp_recog = face.createLBPHFaceRecognizer()
lbp_recog.load(os.path.join(models_dir, 'lbp.yml'))
lbp_label = lbp_recog.predict(img)
print('LBP done')
print(eigen_label, fischer_label, lbp_label)
return eigen_label, fischer_label, lbp_label
def recognizemany(img_file,
detector_xml_path,
models_dir,
dest_img_file,
eigen=True,
fischer=True,
lbp=True,
equalize_hist=False):
img = cv2.imread(img_file)
gray_img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
#gray_img = cv2.resize(gray_img, (640, 480))
detector = cv2.CascadeClassifier(detector_xml_path)
min_size = (min(50,
gray_img.shape[0] // 10), min(50, gray_img.shape[1] // 10))
#min_size = (0,0)
hits = detector.detectMultiScale(gray_img, 1.1, 3, 0, min_size)
eigen_label = fischer_label = lbp_label = -1
with open(os.path.join(models_dir, 'model.json'), 'r') as model_file:
model = json.load(model_file)
train_img_size = (model['height'], model['width'])
labels = model['labels']
print('# hits:', len(hits))
hits_img = np.copy(img)
# If training images were equalized, better to perform the same
# operation during recognition too.
if equalize_hist:
gray_img = cv2.equalizeHist(gray_img)
i = 1
for (x, y, w, h) in hits:
print('ROI ', i)
roi = gray_img[y:y + h, x:x + w]
i += 1
if roi.shape != train_img_size:
roi = cv2.resize(roi, train_img_size[::-1])
if eigen:
eigen_recog = face.createEigenFaceRecognizer()
eigen_recog.load(os.path.join(models_dir, 'eigen.yml'))
eigen_label = eigen_recog.predict(roi)
print('Eigen done')
if fischer:
fischer_recog = face.createFisherFaceRecognizer()
fischer_recog.load(os.path.join(models_dir, 'fischer.yml'))
fischer_label = fischer_recog.predict(roi)
print('Fischer done')
if lbp:
lbp_recog = face.createLBPHFaceRecognizer()
lbp_recog.load(os.path.join(models_dir, 'lbp.yml'))
lbp_label = lbp_recog.predict(roi)
print('LBP done')
cv2.rectangle(hits_img, (x, y), (x + w, y + h), (255, 255, 255), 2)
cv2.putText(hits_img, labels[str(fischer_label)], (x, y - 5),
cv2.FONT_HERSHEY_PLAIN, 2.0, (255, 255, 255), 2)
print(labels[str(eigen_label)], labels[str(fischer_label)],
labels[str(lbp_label)])
#return eigen_label, fischer_label, lbp_label
cv2.imwrite(dest_img_file, hits_img)
def test(test_csv, models_dir, eigen=True, fischer=True, lbp=True):
eigen_label = fischer_label = lbp_label = -1
if eigen:
eigen_recog = face.createEigenFaceRecognizer()
eigen_recog.load(os.path.join(models_dir, 'eigen.yml'))
if fischer:
fischer_recog = face.createFisherFaceRecognizer()
fischer_recog.load(os.path.join(models_dir, 'fischer.yml'))
if lbp:
lbp_recog = face.createLBPHFaceRecognizer()
lbp_recog.load(os.path.join(models_dir, 'lbp.yml'))
with open(os.path.join(models_dir, 'model.json'), 'r') as model_file:
train_img_size = json.load(model_file)
train_img_size = (train_img_size['height'], train_img_size['width'])
#test_imgfiles = np.genfromtxt(test_csv, delimiter=',', dtype=None, names=['file','label','labelnum'])
test_imgfiles = []
with open(test_csv, 'r', encoding='utf-8', newline='') as csvfile:
reader = csv.reader(csvfile)
for row in reader:
f, label, label_idx = row
test_imgfiles.append(row)
eigen_correct = fischer_correct = lbph_correct = 0
for test_imgfile in test_imgfiles:
img = cv2.imread(test_imgfile[0], cv2.IMREAD_GRAYSCALE)
if img.shape != train_img_size:
img = cv2.resize(img, train_img_size[::-1])
expected_label = int(test_imgfile[2])
eigen_label, eigen_conf = eigen_recog.predict(img) if eigen else (-1,
0)
eigen_correct += 1 if eigen_label == expected_label else 0
fischer_label, fischer_conf = fischer_recog.predict(
img) if fischer else (-1, 0)
fischer_correct += 1 if fischer_label == expected_label else 0
lbp_label, lbp_conf = lbp_recog.predict(img) if lbp else (-1, 0)
lbph_correct += 1 if lbp_label == expected_label else 0
print("%s: expected=%d | eigen=%d | fischer=%d | lbph=%d\n" %
(test_imgfile[0], expected_label, eigen_label, fischer_label,
lbp_label))
if eigen:
print("Eigenfaces accuracy: ", eigen_correct / len(test_imgfiles))
if fischer:
print("Fischerfaces accuracy: ", fischer_correct / len(test_imgfiles))
if lbp:
print("LBPH accuracy: ", lbph_correct / len(test_imgfiles))
def train(csv_file,
train_percent,
test_file_csv,
models_dir,
eigen=True,
fischer=True,
lbp=True):
# OMG np.genfromtxt is horribly broken when moving from py2 to py3 because it returns byte arrays in py3
# and nothing else can handle byte arrays properly without other conversion hacks.
# Whatever happened to the "pythonic" way?! Avoid!
# data = np.genfromtxt(csv_file, delimiter=',', dtype=None, names=['file','label','labelnum'])
data = []
all_labels = {}
label_counts = {}
labelnum_col = []
with open(csv_file, 'r', encoding='utf-8', newline='') as csvfile:
reader = csv.reader(csvfile)
for row in reader:
f, label, label_idx = row
labelnum_col.append(label_idx)
data.append(row)
if all_labels.get(label_idx) is None:
all_labels[label_idx] = label
label_counts[label_idx] = 1
else:
label_counts[label_idx] += 1
# Every label should have atleast 2 data points. Delete those rows which don't
# satisfy that condition.
data = [d for d in data if label_counts[d[2]] >= 2]
labelnum_col = [d[2] for d in data]
train_imagefiles, test_imagefiles = train_test_split(
data, train_size=train_percent / 100.0, stratify=labelnum_col)
with open(test_file_csv, 'w', encoding='utf-8') as csvfile:
testwriter = csv.writer(csvfile, delimiter=',')
for test_imgfile in test_imagefiles:
testwriter.writerow(list(test_imgfile))
training_labels = np.array([d[2] for d in train_imagefiles],
dtype=np.int32)
train_images = []
for train_imgfile in train_imagefiles:
#f = train_imgfile[0].decode("utf-8")
print(f)
img = cv2.imread(train_imgfile[0], cv2.IMREAD_GRAYSCALE)
print(f, img.shape)
train_images.append(img)
if not os.path.exists(models_dir):
os.makedirs(models_dir)
print(train_images[0].shape, len(training_labels))
if eigen:
eigen_recog = face.createEigenFaceRecognizer()
eigen_recog.train(train_images, training_labels)
eigen_recog.save(os.path.join(models_dir, 'eigen.yml'))
print('Eigen done')
if fischer:
fischer_recog = face.createFisherFaceRecognizer()
fischer_recog.train(train_images, training_labels)
fischer_recog.save(os.path.join(models_dir, 'fischer.yml'))
print('Fischer done')
if lbp:
lbp_recog = face.createLBPHFaceRecognizer()
lbp_recog.train(train_images, training_labels)
lbp_recog.save(os.path.join(models_dir, 'lbp.yml'))
print('LBP done')
# Record the training image dimensions because at prediction time we need to resize images
# to those dimensions.
model = {
'width': train_images[0].shape[1],
'height': train_images[0].shape[0],
'labels': all_labels
}
with open(os.path.join(models_dir, 'model.json'), 'w') as model_file:
json.dump(model, model_file, indent=4, separators=(',', ': '))
#!/usr/bin/env python
import argparse
import cv2
import cv2.face as face
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Recognize an image against a trained data set (currently EigenFaceRecognizer)')
parser.add_argument('training_filename', help='The training data generated by 2.train.py')
parser.add_argument('image_filename', help='The image containing the face to recognize')
parser.add_argument('-n', '--num-components', type=int, default=10, help='Number of components used in training (default: 10)')
parser.add_argument('-t', '--threshold', type=float, default=12.5, help='Confidence threshold (default: 12.5)')
args = parser.parse_args()
model = face.createEigenFaceRecognizer(args.num_components, args.threshold)
model.load(args.training_filename)
# Load the image to recognize. Must be greyscale
predict_img = cv2.imread(args.image_filename, cv2.IMREAD_GRAYSCALE)
label, confidence = model.predict(predict_img)
print("The predicted label was: %d" % label)
print("The confidence was: %f" % confidence)
testingdict[csvrow['path']] = (img, int(csvrow['id']))
elif ratio:
## Separate some images, to be used for testing. Do not train on them.
testsize = int(math.floor(ratio*TOTAL_IMAGES))
for i in range(testsize):
randpath = random.choice(trainingdict.keys())
randimage = trainingdict[randpath][0]
randid = trainingdict[randpath][1]
testingdict[randpath] = (randimage, randid)
trainingdict.pop(randpath, None)
## Train Eigenfaces, by feeding it face data
trainingimgs, trainingids = zip(*trainingdict.values())
testingimgs, testingids = zip(*testingdict.values())
model = cv2face.createEigenFaceRecognizer()
model.train(numpy.array(trainingimgs), numpy.array(trainingids))
## Test our model and see how accurate we are
correct = 0
iscorrect = False
for path,x in testingdict.iteritems():
img = x[0]
actualid = x[1]
predictionid = model.predict(numpy.array(img))
if predictionid == actualid: # We got it!
correct += 1
iscorrect = True
else: # Dang it!
predictionpath = "" # BUG: If no id is found, cause they all went into testing, this will fail.
for path2,x2 in trainingdict.iteritems():
