def __init__(self, cascade, mdl_landmarks, proto_caffe, mdl_caffe,
blob_name, mdl_start_pth):
self.cascade = cv2.CascadeClassifier(cascade)
self.alignment = ADlib(mdl_landmarks)
self.net = caffe.Net(proto_caffe, mdl_caffe, caffe.TEST)
self.blob_name = blob_name
self.classify = svm_load_model(mdl_start_pth)
def __init__(self, cascade, mdl_landmarks, proto_caffe, mdl_caffe, blob_name, mdl_start_pth):
self.cascade = cv2.CascadeClassifier(cascade)
self.alignment = ADlib(mdl_landmarks)
self.net = caffe.Net(proto_caffe, mdl_caffe, caffe.TEST)
self.blob_name = blob_name
self.classify = svm_load_model(mdl_start_pth)
class faceRecModel:
def __init__(self, cascade, mdl_landmarks, proto_caffe, mdl_caffe, blob_name, mdl_start_pth):
self.cascade = cv2.CascadeClassifier(cascade)
self.alignment = ADlib(mdl_landmarks)
self.net = caffe.Net(proto_caffe, mdl_caffe, caffe.TEST)
self.blob_name = blob_name
self.classify = svm_load_model(mdl_start_pth)
def faceDetCV(self, frm, skinFilter=True):
bbs = self.cascade.detectMultiScale(cv2.equalizeHist(cv2.cvtColor(frm, cv2.COLOR_BGR2GRAY)), scaleFactor=1.1, minNeighbors=3, minSize=(40, 40), flags=cv2.cv.CV_HAAR_SCALE_IMAGE)
bbs = bbs if len(bbs) else np.array([])
if len(bbs) and skinFilter:
keep = np.array([False] * bbs.shape[0])
i = 0
for (x,y,w,h) in bbs:
bb_hsv = cv2.cvtColor(frm[y:y+h, x:x+w], cv2.COLOR_BGR2HSV)
skinMask = cv2.inRange(bb_hsv, lower, upper)
skinMask = cv2.dilate(skinMask, kernel, iterations=2)
skinMask = skinMask[h/5 : h, w/4 : 3*w/4]
if np.sum(skinMask) * 2 / (h * w * 255 ) > 0.4:
keep[i] = True
i += 1
return bbs[keep]
else:
return bbs
def faceDetAlign(self, frm):
bbs = self.faceDetCV(frm, skinFilter=True)
print "detected ", len(bbs), " faces"
faces = None
scale = max(frm.shape[:2]) / 960.0
frm_small = cv2.resize(frm, (int(frm.shape[1] / scale), int(frm.shape[0] / scale)), interpolation=cv2.INTER_AREA)
frm_small = cv2.cvtColor(frm_small, cv2.COLOR_BGR2RGB)
bbs_small = (bbs / scale).astype('int')
xmax = frm_small.shape[1]
ymax = frm_small.shape[0]
bbs_filt = []
offset = 15
for (x,y,w,h) in bbs_small:
x0 = max(x-offset, 0)
y0 = max(y-offset, 0)
x1 = min(x+w+2*offset, xmax-1)
y1 = min(y+h+2*offset, ymax-1)
bb = self.alignment.getLargestFaceBoundingBox(np.array(frm_small[y0:y1, x0:x1]))
if bb is None:
print 'CV false positive'
continue
bb = dlib.rectangle(int((bb.left()+x0)*scale), int((bb.top()+y0)*scale), int((bb.right()+x0)*scale), int((bb.bottom()+y0)*scale))
bbs_filt.append(bb)
landmarks = np.float32(self.alignment.findLandmarks(frm, bb))
H = cv2.getAffineTransform(np.vstack((landmarks[[36, 45]], (landmarks[48]+landmarks[54])/2.)), np.vstack((Anchors[[0, 1]], (Anchors[2]+Anchors[3])/2.)))
face_aligned = cv2.warpAffine(frm, H, (128, 128))
face_aligned = cv2.cvtColor(face_aligned, cv2.COLOR_RGB2GRAY)
face_aligned = cv2.equalizeHist(face_aligned)
face_aligned = face_aligned[np.newaxis, np.newaxis, :, :] / 255.0
faces = face_aligned if faces is None else np.vstack((faces, face_aligned))
return faces, bbs_filt
def featureExtract(self, faces):
if faces is None:
return []
self.net.blobs['data'].reshape(*faces.shape)
self.net.blobs['data'].data[...] = faces
self.net.forward()
feats = self.net.blobs[self.blob_name].data
return feats
def runSVM(self, feature, model):
p_label, p_acc, p_vals = svm_predict([0] * len(feature), feature, model, '-b 1 -q')
return p_label, p_vals
def serv(self, inp_q, res_q, continue_evt, reload_val, mdl_update_pth):
while True:
try:
continue_evt.wait()
if reload_val.value:
self.classify = svm_load_model(mdl_update_pth)
reload_val.value = 0
cli_name, data_buffer, mode = inp_q.get(timeout=2)
bbs_filt, labels, probs = [], [], []
if mode: # 1 is strong mode
bbs_filt_np, feats = data_buffer
bbs_filt_np = bbs_filt_np.astype(np.int64)
for bbs_np in bbs_filt_np:
bbs_filt.append(dlib.rectangle(bbs_np[0], bbs_np[1], bbs_np[2], bbs_np[3]))
else:
nparr = np.fromstring(data_buffer, dtype=np.uint8)
img_np = cv2.imdecode(nparr, cv2.CV_LOAD_IMAGE_COLOR)
faces, bbs_filt = self.faceDetAlign(img_np)
feats = self.featureExtract(faces)
if len(feats):
labels, probs = self.runSVM(feats.tolist(), self.classify)
res_q.put((cli_name, (bbs_filt, labels, probs, feats)))
inp_q.task_done()
except (Queue.Empty, KeyboardInterrupt):
# print 'timeout in face'
pass
class faceRecModel:
def __init__(self, cascade, mdl_landmarks, proto_caffe, mdl_caffe,
blob_name, mdl_start_pth):
self.cascade = cv2.CascadeClassifier(cascade)
self.alignment = ADlib(mdl_landmarks)
self.net = caffe.Net(proto_caffe, mdl_caffe, caffe.TEST)
self.blob_name = blob_name
self.classify = svm_load_model(mdl_start_pth)
def faceDetCV(self, frm, skinFilter=True):
bbs = self.cascade.detectMultiScale(cv2.equalizeHist(
cv2.cvtColor(frm, cv2.COLOR_BGR2GRAY)),
scaleFactor=1.1,
minNeighbors=3,
minSize=(40, 40),
flags=cv2.cv.CV_HAAR_SCALE_IMAGE)
bbs = bbs if len(bbs) else np.array([])
if len(bbs) and skinFilter:
keep = np.array([False] * bbs.shape[0])
i = 0
for (x, y, w, h) in bbs:
bb_hsv = cv2.cvtColor(frm[y:y + h, x:x + w], cv2.COLOR_BGR2HSV)
skinMask = cv2.inRange(bb_hsv, lower, upper)
skinMask = cv2.dilate(skinMask, kernel, iterations=2)
skinMask = skinMask[h / 5:h, w / 4:3 * w / 4]
if np.sum(skinMask) * 2 / (h * w * 255) > 0.4:
keep[i] = True
i += 1
return bbs[keep]
else:
return bbs
def faceDetAlign(self, frm):
bbs = self.faceDetCV(frm, skinFilter=True)
print "detected ", len(bbs), " faces"
faces = None
scale = max(frm.shape[:2]) / 960.0
frm_small = cv2.resize(
frm, (int(frm.shape[1] / scale), int(frm.shape[0] / scale)),
interpolation=cv2.INTER_AREA)
frm_small = cv2.cvtColor(frm_small, cv2.COLOR_BGR2RGB)
bbs_small = (bbs / scale).astype('int')
xmax = frm_small.shape[1]
ymax = frm_small.shape[0]
bbs_filt = []
offset = 15
for (x, y, w, h) in bbs_small:
x0 = max(x - offset, 0)
y0 = max(y - offset, 0)
x1 = min(x + w + 2 * offset, xmax - 1)
y1 = min(y + h + 2 * offset, ymax - 1)
bb = self.alignment.getLargestFaceBoundingBox(
np.array(frm_small[y0:y1, x0:x1]))
if bb is None:
print 'CV false positive'
continue
bb = dlib.rectangle(int((bb.left() + x0) * scale),
int((bb.top() + y0) * scale),
int((bb.right() + x0) * scale),
int((bb.bottom() + y0) * scale))
bbs_filt.append(bb)
landmarks = np.float32(self.alignment.findLandmarks(frm, bb))
H = cv2.getAffineTransform(
np.vstack(
(landmarks[[36,
45]], (landmarks[48] + landmarks[54]) / 2.)),
np.vstack((Anchors[[0, 1]], (Anchors[2] + Anchors[3]) / 2.)))
face_aligned = cv2.warpAffine(frm, H, (128, 128))
face_aligned = cv2.cvtColor(face_aligned, cv2.COLOR_RGB2GRAY)
face_aligned = cv2.equalizeHist(face_aligned)
face_aligned = face_aligned[np.newaxis, np.newaxis, :, :] / 255.0
faces = face_aligned if faces is None else np.vstack(
(faces, face_aligned))
return faces, bbs_filt
def featureExtract(self, faces):
if faces is None:
return []
self.net.blobs['data'].reshape(*faces.shape)
self.net.blobs['data'].data[...] = faces
self.net.forward()
feats = self.net.blobs[self.blob_name].data
return feats
def runSVM(self, feature, model):
p_label, p_acc, p_vals = svm_predict([0] * len(feature), feature,
model, '-b 1 -q')
return p_label, p_vals
def serv(self, inp_q, res_q, continue_evt, reload_val, mdl_update_pth):
while True:
try:
continue_evt.wait()
if reload_val.value:
self.classify = svm_load_model(mdl_update_pth)
reload_val.value = 0
cli_name, data_buffer, mode = inp_q.get(timeout=2)
bbs_filt, labels, probs = [], [], []
if mode: # 1 is strong mode
bbs_filt_np, feats = data_buffer
bbs_filt_np = bbs_filt_np.astype(np.int64)
for bbs_np in bbs_filt_np:
bbs_filt.append(
dlib.rectangle(bbs_np[0], bbs_np[1], bbs_np[2],
bbs_np[3]))
else:
nparr = np.fromstring(data_buffer, dtype=np.uint8)
img_np = cv2.imdecode(nparr, cv2.CV_LOAD_IMAGE_COLOR)
faces, bbs_filt = self.faceDetAlign(img_np)
feats = self.featureExtract(faces)
if len(feats):
labels, probs = self.runSVM(feats.tolist(), self.classify)
res_q.put((cli_name, (bbs_filt, labels, probs, feats)))
inp_q.task_done()
except (Queue.Empty, KeyboardInterrupt):
# print 'timeout in face'
pass
