def appearance(mean,points,training):
xmean = mean[:len(mean)/2]
ymean = mean[len(mean)/2:]
minxmean = np.min(xmean)
minymean = np.min(ymean)
refshape = zip(xmean,ymean)
refshape = [((p[0]-minxmean),(p[1]-minymean)) for p in refshape]
model_matrix = []
appshapes = np.zeros(shape=(len(training),2))
# get vector for each image
for i in range(len(training)):
# take gradient + clean
gradimg = rg.togradient(training[i])
# get model from cleaned image
appearance = appvec(refshape,points[i],gradimg)
model_matrix.append(appearance)
appshapes[i] = appearance.shape
#calculate mean appearance shape
meanshape0 = int(np.mean([s[0] for s in appshapes]))
meanshape1 = int(np.mean([s[1] for s in appshapes]))
model = np.uint8(np.zeros(shape=(len(training),meanshape0*meanshape1)))
#resize model matrix according to mean shape and flatten into vector
for i in range(len(training)):
model[i] = cv2.resize(model_matrix[i],(meanshape1,meanshape0)).flatten()
# mean of image model vectors and mean of shape = total model
return np.uint8(np.mean(model,axis=0)),(meanshape0, meanshape1)
def autoinit(mean, marks, trainimgs, testimg, k, m, gmod):
imgw, imgh = testimg.shape[1],testimg.shape[0]
# get appearance model
a_mean,mean_shape = appearance(mean,marks,trainimgs)
# get TL model (mean and standard dev) for top-left corners of landmarks
TL, TLmin, TLmax = TLmodel(marks,trainimgs, imgw, imgh)
# get radiograph as gradient image (also equalizes for better edges)
gradimg = rg.togradient(testimg)
#
xmean, ymean = mean[:len(mean)/2], mean[len(mean)/2:]
min_x,min_y = np.min(xmean),np.min(ymean)
initshape = zip([(x-min_x)*imgw for x in xmean],[(y-min_y)*imgh for y in ymean])
min_x,min_y = np.min(mean[:len(mean)/2]),np.min(mean[len(mean)/2:])
pts = zip(mean[:len(mean)/2],mean[len(mean)/2:])
pts = [((point[0]-min_x),(point[1]-min_y)) for point in pts]
initshape = [(int(y[0]*imgw),int(y[1]*imgh)) for y in pts]
# set of angles and scales to nudge new shapes in
angles = np.array(range(-8,9,1))/2.0
scales = range(8,-1,-1)
# to start, use template equal to bounding box around mean shape of appearance model
template = a_mean.reshape((mean_shape[0],mean_shape[1]))
temph, tempw = template.shape
chosenangle = 0
chosenscalex = 0
chosenscaley = 0
chosenlocation = (0,0)
dmin = np.inf
# scale shape over select amount of scaling factors- x axis
for sx in [0.05*s+0.1 for s in scales]:
xtempl = cv2.resize(template,(int(tempw*sx),temph))
# scale shape over select amount of scaling factors - y axis
for sy in [0.05*s+0.1 for s in scales]:
X,Y,dTL = gettemplateparam(xtempl,sy, TLmin, TLmax, TL, gradimg)
# rotate shape over select amount of rotations, find best fit
for a in angles:
angle = a*0.1
# get a new shape by 'nudging' the current shape
# = scale, rotate, translate
nudgedshape = nudgeshape(sx,sy,angle,X,Y,initshape,1.0,1.0)
points = np.array(zip(*nudgedshape)).flatten().tolist()
d = 0
if not(np.max(points[:len(points)/2])+k > imgw-1 or np.max(points[len(points)/2:])+k > imgh-1):
for l in range(len(nudgedshape)):
# get model, gradients, normal and profile for point l
profile, n = prfl.get(nudgedshape,l)
cov,mean = gmod[l]
# Get mahalanobis distance from profile[0]'s greylvl model to provided greymodel (co,mean)
d = d + prfl.greylvldistancemetric(profile[0], mean, cov, k ,n,testimg)
# calculate full metric
d = d/len(nudgedshape)
dtot = d+0.01*np.linalg.norm(dTL)
# update
if dtot < dmin:
dmin = dtot
chosenlocation = X,Y
chosenscalex = sx
chosenscaley = sy
chosenangle = angle
print "Estimate found:"
print "Location = ("+ ",".join([str(c).split(".")[0] for c in chosenlocation]) + ")"
print "Angle = " + str(chosenangle)
print "Scale X = " + str(chosenscalex)
print "Scale Y = " + str(chosenscaley)
# nudge initial shape by chosen angle, scale and position
result = nudgeshape(chosenscalex,chosenscaley,chosenangle,chosenlocation[0],chosenlocation[1],initshape,(1.0/float(imgw)),(1.0/float(imgh)))
return [el for point in zip(*result) for el in point]
