def preprocess(feature_abstract_method):
# X_raw = raw_data.iloc[:, 1:]
# y_raw = raw_data['label']
# X_train, X_test, y_train, y_test = train_test_split(X_raw, y_raw, test_size=0.2)
# X_train.to_csv('x_train.csv')
# X_test.to_csv('x_test.csv')
# y_train.to_csv('y_train.csv')
# y_test.to_csv('y_test.csv')
X_train = pd.read_csv('x_train.csv', index_col=0)
X_test = pd.read_csv('x_test.csv', index_col=0)
y_train = pd.read_csv('y_train.csv', index_col=0, header=None)
y_test = pd.read_csv('y_test.csv', index_col=0, header=None)
if (feature_abstract_method == 'LBP'):
X_train = LBP.lbp_extract(X_train)
X_test = LBP.lbp_extract(X_test)
elif (feature_abstract_method == 'PCA'):
X_train, X_test = PCA.PCA_extract(X_train, X_test)
elif (feature_abstract_method == 'skeleton'):
X_train = SKELETON.skeleton_extract(X_train)
X_test = SKELETON.skeleton_extract(X_test)
elif (feature_abstract_method == 'grid'):
X_train = GRID.grid_extract(X_train)
X_test = GRID.grid_extract(X_test)
elif (feature_abstract_method == 'hog'):
X_train = HOG.hog_extract(X_train)
X_test = HOG.hog_extract(X_test)
return X_train, X_test, y_train, y_test
def cargar_datos(PATH_POSITIVE, PATH_NEGATIVE):
data = []
clases = []
# Casos positivos
counter_positive_samples = 0
for filename in os.listdir(PATH_POSITIVE):
if filename.endswith(IMAGE_EXTENSION):
filename = PATH_POSITIVE+filename
img = cv2.imread(filename)
hog = cv2.HOGDescriptor()
descriptor_1 = hog.compute(img)
lbp = LBP.LocalBinaryPattern(img)
descriptor_2 = lbp.compute_lbp_clasic()
#descriptor = lbp.compute_lbp_uniform()
descriptor = np.concatenate((descriptor_1, descriptor_2))
data.append(descriptor)
clases.append(1)
counter_positive_samples += 1
print("Leidas " + str(counter_positive_samples) + " imágenes de -> peatones")
print("Leidas " + str(counter_positive_samples) + " imágenes de -> peatones")
print(np.shape(data))
# Casos negativos
counter_negative_samples = 0
for filename in os.listdir(PATH_NEGATIVE):
if filename.endswith(IMAGE_EXTENSION):
filename = PATH_NEGATIVE+filename
img = cv2.imread(filename)
hog = cv2.HOGDescriptor()
descriptor_1 = hog.compute(img)
lbp = LBP.LocalBinaryPattern(img)
descriptor_2 = lbp.compute_lbp_clasic()
#descriptor = lbp.compute_lbp_uniform()
descriptor = np.concatenate((descriptor_1, descriptor_2))
data.append(descriptor)
clases.append(0)
counter_negative_samples += 1
print("Leidas " + str(counter_negative_samples) + " imágenes de -> fondo")
print("Leidas " + str(counter_negative_samples) + " imágenes de -> fondo")
print(np.shape(data))
return np.array(data), np.array(clases)
def recognize(cf, img):
# caricamento dei vari dataset
gallery_target = np.load("npy_db/gallery_target.npy")
histogram_gallery_data = np.load("npy_db/histogram_gallery_data.npy")
gallery_thresholds = np.load("npy_db/gallery_thresholds.npy")
users = pd.read_csv('dataset_user.csv', index_col=[0])
cf_list = users['Codice Fiscale']
# controlla se l'utente esiste
if not cf_list.tolist().__contains__(cf):
print("UTENTE NON PRESENTE! Il codice fiscale inserito è:", cf)
return None, 0, None
# inizializzazione delle variabili che conterranno i dati del paziente e l'indice in cui si trova in dataset_user
user = None
index = cf_list.tolist().index(cf)
# l'immagine in input viene normalizzata per poter utilizzare LBP e ottenerne l'istogramma relativo
norm_image = cv2.normalize(img, None, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F)
norm_image = norm_image.astype(np.uint8)
lbp = LBP.Local_Binary_Pattern(1, 8, norm_image)
hist = lbp.createHistogram(lbp.compute_lbp())
# calcola la corrispondenza massima tra l'istogramma dell'immagine in input e quelli delle immagini nella galleria dell'utente
val = topMatch(cf, gallery_target, histogram_gallery_data, hist)
# se la similairtà massima è maggiore o uguale alla soglia adattativa del paziente, l'identità viene verificata
if val >= gallery_thresholds[index]:
user = users.iloc[index]
# ritorna le informazioni del paziente e l'indice in cui si trova in dataset_user
return user, index, user
def descriptor(name, arg):
model = None
if name == "LBP":
model = LBP(arg)
elif name == "RAD":
model = RAD(arg)
return model
def identify(cf, img):
# caricamento dei vari dataset
gallery_target = np.load("npy_db/gallery_target.npy")
histogram_gallery_data = np.load("npy_db/histogram_gallery_data.npy")
users = pd.read_csv('dataset_user.csv', index_col=[0])
gallery_thresholds = np.load("npy_db/gallery_thresholds.npy")
galley_users = list(dict.fromkeys(gallery_target))
cf_list = users['Codice Fiscale']
# controlla se l'utente esiste
if not cf_list.tolist().__contains__(cf):
print("UTENTE NON PRESENTE! Il codice fiscale inserito è:", cf)
return None, 0, None
# informazioni del paziente utilizzando cf
index = cf_list.tolist().index(cf)
user = users.iloc[index]
# lista dei delegati del paziente
delegati = ast.literal_eval(user["Delegati"])
# se il paziente non ha nessun delegato, allora termina l'operazione
if len(delegati) == 0:
print("L'utente non ha delegati!")
return None, 0, None
# si inizializzano le variabili per la similarity e l'identità del delegato
max = 0
identity = None
# l'immagine in input viene normalizzata per poter utilizzare LBP e ottenerne l'istogramma relativo
norm_image = cv2.normalize(img, None, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F)
norm_image = norm_image.astype(np.uint8)
lbp = LBP.Local_Binary_Pattern(1, 8, norm_image)
hist = lbp.createHistogram(lbp.compute_lbp())
# per ogni delegato
for d in delegati:
# ottengo il miglior valore ottenuto confrontando l'istograma con quelli del delegato nella galleria
val = topMatch(d, gallery_target, histogram_gallery_data, hist)
th_index = galley_users.index(d)
# confrontiamo la similarity con quella massimo ottenuta finora e con il threshold del delegato
# se la similarity supera quella massima e il threshold, allora aggiorniamo le variabili
if val > max and val >= gallery_thresholds[th_index]:
max = val # il più alto valore di similarity attuale
identity = d # identità del delegato che ha ottenuto il miglior valore per il momento
# se c'è stato un riconoscimento tra i delegati
if identity is not None:
indexd = cf_list.tolist().index(identity)
recUser = users.iloc[indexd]
# ritorna i dati del paziente, l'indice in cui si trova in dataset_user e le informazioni del delegato
return user, index, recUser
else:
# altrimenti ritorna None come paziente, 0 come indice e None come delegato
return None, 0, None
def convert_image_to_hist(self, image):
print(image)
image = cv2.imread(image)
norm_image = get_normalized(image)
myLBP = LBP.Local_Binary_Pattern(1, 8, norm_image)
new_img = myLBP.compute_lbp()
hist = myLBP.createHistogram(new_img)
return hist
def split_gallery_probe(self, data, target, cfs):
num_user = self.num_user(target)
# calcola il numero di template per utente
unique, counts = np.unique(target, return_counts=True)
occurrences = dict(zip(unique, counts))
# numero di utenti che non sono nella gallery
pn_user = round(num_user * pn / 100)
# conteggio dei template
countTemp = 0
# conteggio
countUser = 0
# gallery_target, gallery_data, pn_data, pn_target, pg_data, pg_target = [], [], [], [], [], []
gallery_target, gallery_data, pn_data, pn_target, pg_data, pg_target, histogram_gallery_data, \
histogram_pg_data, histogram_pn_data = [], [], [], [], [], [], [], [], []
# prendo il numero di template dell'utente. si parte dal presupposto che tutti gli utenti
# hanno lo stesso numero di template
occ = occurrences[target[0]]
# numero di template per il probe set
self.n_template_x_user = round(occ * probe / 100)
# per ogni utente
for i, val in enumerate(target):
name = cfs[countUser]
norm_template = self.get_normalized_template(i, data)
lbp = LBP.Local_Binary_Pattern(1, 8, norm_template)
# se il numero del template e' minore del numero massimo di template destinati alla gallery
# e il numero di utenti rientra negli utenti che sono nella gallery, allora inserisco il template nella
# gallery
if (countTemp < occ - self.n_template_x_user
or occ == 1) and countUser < num_user - pn_user:
gallery_data.append(norm_template)
gallery_target.append(name)
histogram_gallery_data.append(
lbp.createHistogram(lbp.compute_lbp()))
else:
# se lo inserisco nel probe set, controllo se l'utente e' nella gallery o no
if countUser < num_user - pn_user:
pg_data.append(norm_template)
pg_target.append(name)
histogram_pg_data.append(
lbp.createHistogram(lbp.compute_lbp()))
else:
pn_data.append(norm_template)
pn_target.append(name)
histogram_pn_data.append(
lbp.createHistogram(lbp.compute_lbp()))
countTemp += 1
# se ho finito i template dell'utente, passo all'utente successivo
if countTemp == occ:
countTemp = 0
countUser += 1
return gallery_data, gallery_target, pn_data, pn_target, pg_data, pg_target, histogram_gallery_data, histogram_pg_data, histogram_pn_data
def __init__(self, model_name, folder_path=None):
self.model_name = model_name
self.folder_path = folder_path
self.split_windows = False
self.model = None
if self.model_name == 'LBP':
self.model = LBP.LocalBinaryPatterns(8, 1)
self.split_windows = True
elif self.model_name == 'HOG':
self.model = HOG.Hog(orientations=9, pixels_per_cell=(8, 8), cells_per_block=(2, 2))
elif self.model_name == 'CM':
self.model = ColorMoments.ColorMoments()
self.split_windows = True
elif self.model_name == 'SIFT':
self.model = SIFT.SIFT()
def microTextureVideo(self, pathVid):
cap = cv2.VideoCapture(pathVid)
val = False
# andiamo a prendere un frame e lo convertiamo in scala di grigi.
while True:
ret, frame = cap.read()
try:
vis = frame.copy()
except Exception as e:
print(str(e))
break
# convertiamo in scala di grigi.
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
# effettuiamo il ritaglio del viso
crop = detect_face(gray, vis)
# qui effettuiamo la conversione dell'immagine croppata in LBP
if crop is not None:
myLBP = LBP.Local_Binary_Pattern(1, 8, crop)
else:
continue
new_img = myLBP.compute_lbp()
# creiamo l'histogram dell'immagine calcolata in lpb
hist = myLBP.createHistogram(new_img)
# Andiamo a prendere il modello trained e salvato.
with open('modelSVM.pkl', 'rb') as f:
clf = pickle.load(f)
hist = hist.reshape(1, -1)
# attraverso il classificatore che abbiamo recuperato, ci facciamo dire se l'immagine è reale oppure no
value = (clf.predict(hist))
print(value)
if value == 0:
print("REAL")
val = True
break
else:
print("FAKE")
val = False
break
cap.release()
cv2.destroyAllWindows()
return val
def main():
ap = argparse.ArgumentParser(
description='Run the local binary patterns algorithm using a basic 3x3.'
)
ap.add_argument('-i',
'--input',
dest='input',
type=str,
required=True,
help='file name with path of the input image')
arguments = ap.parse_args()
input_file = arguments.input #'data/simpsons/Test/bart116.jpg'
if os.path.isfile(input_file):
run = LBP.LBP(input_file)
print("RUNNING algorithm developed")
run.execute()
#print("RUNNING scikit-image")
#run.compare()
else:
print("File '{}' does not exist.".format(input_file))
'C:/Users/rebeb/Documents/TU_Wien/Dipl/FID-300/FID-300/FID-300/test_images/results/SIFT/output_SIFT_matched_00197_alternative.jpg',
outputMatched)
cv.imshow("keypoint mask", results)
cv.imshow("matched", outputMatched)
cv.waitKey(0)
cv.destroyAllWindows()
if LBPLearning:
# patterns = LBP.threeLayeredLearning(images, masks)
patterns = np.loadtxt(
'C:/Users/rebeb/Documents/TU_Wien/Dipl/FID-300/FID-300/FID-300/test_images/training/discriminative_4_12_3.txt',
delimiter=',')
img = cv.imread(
'C:/Users/rebeb/Documents/TU_Wien/Dipl/FID-300/FID-300/FID-300/test_images/hard/00197.jpg',
0)
lbpImage = LBP.getLBPImage(img, 4, 12, 3)
height, width = img.shape
res = np.zeros((height, width), np.float32)
print("calculating")
for x in range(width):
for y in range(height):
currentHisogram = lbpImage[y, x]
currentHisogram = np.float32(currentHisogram)
maxVal = 0
for pattern in patterns:
val = cv.compareHist(currentHisogram,
np.float32(np.asarray(pattern)),
cv.HISTCMP_CORREL)
# if val > 0.75:
# res[y, x] = 1
# continue
methods = ["HOG","LBP","HOGLBP"]
method = methods[0]
if method == "HOG":
print("__________________________________________________________________")
print("Computing HOG...")
X_train = compute_hog_set(X)
elif method == "LBP":
print("__________________________________________________________________")
print("Computing LBP...")
# Calculate LBP values for each image
lbp_images = []
for i in range(0, len(X)):
lbp_images.append(lbp.lbp_compute(X[i]))
# Calculate histograrms for each image
uniform = False
print("Computing Histograms...uniform:",str(uniform))
histograms = []
for i in range(0, len(lbp_images)):
# IMPORTANT, set uniform to False or True for LBP/LBPU
histograms.append(lbp.lbp_hist(lbp_images[i], step=8, win_size=16,
uniform = uniform))
X = [elem for histogram in histograms for elem in histogram]
for i in range(0,len(histograms)):
histograms[i] = np.concatenate(histograms[i])
test_images, test_classes = utils.load_data(TEST_IMAGES_FOLDER, IMG_EXTENSION)
print("Calculando descriptores")
train_descriptors = utils.compute_lbp(train_images, uniform=True)
test_descriptors = utils.compute_lbp(test_images, uniform=True)
descriptors = np.vstack((train_descriptors, test_descriptors))
labels = np.concatenate((train_classes, test_classes))
print("Entrenando el clasificador")
classifier = utils.train(descriptors.astype(np.float32),
labels,
kernel=cv2.ml.SVM_POLY,
params={'degree': 2})
detector = LBP.LBPDetector(8, 8, 16, 16, 128, 64, 8, 8, classifier)
print("Detectando en Abbey Road")
abbey_road = cv2.imread("abbey_road.jpeg")
peds = detector.detect(abbey_road, [0.8, 1, 1.2])
for ped in peds:
cv2.rectangle(abbey_road, (ped[1], ped[0]), (ped[3], ped[2]), (0, 0, 255),
1)
cv2.imwrite("abbey_road_dets_multiscale.jpeg", abbey_road)
print("Detectando en Pedestrians")
pedestrians = cv2.imread("street.jpg")
histograma = 0
for (i,rect) in enumerate(rects):
if imagini.endswith(".jpg"):
cv2.imshow("imaginegrey",img_data)
(x, y, w, h) = face_utils.rect_to_bb(rect)
b = cv2.rectangle(img_data, (x, y), (x + w, y + h),10,0)
c = b[y:y+h,x:x+w]
try:
c = cv2.resize(c,(154,154))
c1 = copy.deepcopy(c)
lbp = LBP.LBP(c,c1)
cnt_imagini_total += 1
histograma = calcul_histograma.calc_hist(lbp)
l_imag.append(folder)
l_hist_imag.append(histograma)
f = open(subfolder+"_"+imagini[0:-4]+"_"+"hist.txt", "w+")
f.write(str(list(histograma)))
f.close()
lbp = 0
folder = ""
def classifier(name, arg):
model = None
if name == "asdfasdfnlasdflas":
model = LBP(arg)
return model
# =======================
# now for the main script
# =======================
# first set basic display options
numpy.set_printoptions(precision=3)
#numpy.set_printoptions(linewidth=135)
matplotlib.pyplot.ion()
# then run some very basic tests.
# this is to catch silly bugs while developing the library being used.
print("testing basic MRF creation and message passing.")
mrf = LBP.MRF(7, 7, 3)
base_beliefs = numpy.ones(shape=(7, 7, 3))
mrf.init_base_belief(base_beliefs)
smoothness = numpy.ones(shape=(3, 3))
mrf.init_smoothness(smoothness)
mrf.pass_messages()
#print(mrf)
# as that must have worked, now test on the tsukuba pair.
# this will be much more involved,
# but is still a fairly basic application.
# most of the work here is in setting up the base belief array.
print("testing tsukaba pair stereo matching")
input_left = matplotlib.image.imread('tsukuba-imL.png')
def compute_lbp(img_list, uniform=False):
lbp = LBP.LBPDescriptor(8, 8, 16, 16, 128, 64, uniform)
lbp_list = [lbp.compute(img) for img in img_list]
return np.array(lbp_list)
