def extract_features_img(st):
img = get_image(st)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
X_0 = lbp_features(gray, hdiv=1, vdiv=1, mapping='nri_uniform')
XRed = lbp_features(img[:, :, 0], hdiv=1, vdiv=1, mapping='nri_uniform')
XGreen = lbp_features(img[:, :, 1], hdiv=1, vdiv=1, mapping='nri_uniform')
XBlue = lbp_features(img[:, :, 2], hdiv=1, vdiv=1, mapping='nri_uniform')
X_lbp = np.concatenate((X_0, XRed, XGreen, XBlue))
features = np.asarray(X_lbp)
return features
def predict_from_scratch(filepath,
data_filename,
Xtrain=None,
Ytrain=None,
knn=None):
if Xtrain is None or Ytrain is None:
assert knn, "Por favor dar el modelo, o bien los datos Xtrain y Ytrain"
# Obtenemos los valores guardados
with open(data_filename) as f:
data = json.load(f)
hdiv = int(data['hdiv'])
vdiv = int(data['vdiv'])
local_p_clean = np.array(data['p_clean'])
local_p_sfs = np.array(data['p_sfs'])
# Calculamos las features
im = imread(filepath)
lbp = lbp_features(im, hdiv=hdiv, vdiv=vdiv)
# Filtramos
lbp_clean = np.array([lbp[i] for i in local_p_clean])
lbp_sfs = np.array([lbp_clean[i] for i in local_p_sfs])
lbp_sfs = lbp_sfs.reshape(1, -1)
if not knn:
knn = KNeighborsClassifier(n_neighbors=1)
knn.fit(Xtrain, Ytrain)
local_pred = knn.predict(lbp_sfs)
print(f"La predicción es: {local_pred}")
def main(path, n= 50):
#Abrimos el clasificador que entrenamos previamente
with open('T02/trained_classifiers/{}.p'.format(n), 'rb') as fp:
classifier, selected_chars = pickle.load(fp)
_width, _height, rows, _info = png.Reader(filename= path).read()
img_lbp = [lbp_features(get_greyscale_matrix(rows), hdiv=1, vdiv=1)]
Y_test = filter_matrix_by_columns(img_lbp, selected_chars)
return classifier.predict(Y_test)
def create_info(train_0_path= "T02/Training_0/",
train_1_path= "T02/Training_1/",
test_0_path= "T02/Testing_0/",
test_1_path= "T02/Testing_1/"):
train = {0: [], 1: []}
test = {0: [], 1: []}
# All features from train
if not os.path.exists('T02/train_lbps.p'):
for img_name in os.listdir(train_0_path):
if ".png" not in img_name:
continue
_width, _height, rows, _info = png.Reader(filename="{}{}".format(train_0_path, img_name)).read()
train[0].append(lbp_features(get_greyscale_matrix(rows), vdiv=1, hdiv=1))
for img_name in os.listdir(train_1_path):
if ".png" not in img_name:
continue
_width, _height, rows, _info = png.Reader(filename="{}{}".format(train_1_path, img_name)).read()
train[1].append(lbp_features(get_greyscale_matrix(rows), vdiv=1, hdiv=1))
with open('T02/train_lbps.p', 'wb') as fp:
pickle.dump(train, fp)
# All features from test
if not os.path.exists('T02/test_lbps.p'):
for img_name in os.listdir(test_0_path):
if ".png" not in img_name:
continue
_width, _height, rows, _info = png.Reader(filename="{}{}".format(test_0_path, img_name)).read()
test[0].append(lbp_features(get_greyscale_matrix(rows), vdiv=1, hdiv=1))
for img_name in os.listdir(test_1_path):
if ".png" not in img_name:
continue
_width, _height, rows, _info = png.Reader(filename="{}{}".format(test_1_path, img_name)).read()
test[1].append(lbp_features(get_greyscale_matrix(rows), vdiv=1, hdiv=1))
with open('T02/test_lbps.p', 'wb') as fp:
pickle.dump(test, fp)
def extract_features(dirpath, fmt):
st = '*.' + fmt
img_names = dirfiles(dirpath + '/', st)
n = len(img_names)
print(n)
for i in range(n):
img_path = img_names[i]
img = get_image(dirpath + '/' + img_path)
escala_grises = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
X_0 = lbp_features(escala_grises, hdiv=1, vdiv=1, mapping='nri_uniform')
rojo = lbp_features(img[:, :, 0], hdiv=1, vdiv=1, mapping='nri_uniform')
verde = lbp_features(img[:, :, 1], hdiv=1, vdiv=1, mapping='nri_uniform')
azul = lbp_features(img[:, :, 2], hdiv=1, vdiv=1, mapping='nri_uniform')
# Haralick = haralick_features(img.astype(int))
# hog = hog_features(i, v_windows=3, h_windows=3, n_bins=8)
features = np.asarray(np.concatenate((X_0, rojo, verde, azul)))
if i == 0:
m = features.shape[0]
data = np.zeros((n, m))
print('size of extracted features:')
print(features.shape)
data[i] = features
return data
def local_binary_patterns(img, shape, mapping):
pixel_w, pixels_h = shape
return feature_extraction.lbp_features(img,
hdiv=pixel_w,
vdiv=pixels_h,
mapping=mapping)
def extract_features_image(fxi, img=None, bw=None):
# img grayscale image, bw binary image
# check if I2 is a binary image (segmented region)
if fxi == 'lbp':
X = lbp_features(img, region=bw, hdiv=1, vdiv=1, mapping='nri_uniform')
elif fxi == 'lbp-ri':
X = lbp_features(img, region=bw,hdiv=1, vdiv=1, mapping='uniform')
elif fxi == 'gabor':
nr = 8
nd = 8
X = gabor_features(img, region=bw,rotations=nr, dilations=nd)
elif fxi == 'gabor-ri':
nr = 8
nd = 8
Y = gabor_features(img, region=bw,rotations=nr, dilations=nd)
X = np.zeros((nd+3,))
for j in range(nd):
X[j] = np.sum(X[j*nr:(j+1)*nr-1])
X[nr] = Y[nr*nd]
X[nr+1] = Y[nr*nd+1]
X[nr+2] = Y[nr*nd+2]
elif fxi == 'hog':
X = hog_features(img, region=bw, v_windows=1, h_windows=1, n_bins=9,
normalize=False, labels=False, show=False)
elif fxi == 'haralick-1':
X = haralick_features(img, region=bw,distance=1)
elif fxi == 'haralick-2':
X = haralick_features(img, region=bw,distance=2)
elif fxi == 'haralick-3':
X = haralick_features(img, region=bw,distance=3)
elif fxi == 'haralick-5':
X = haralick_features(img, region=bw,distance=5)
elif fxi == 'haralick-7':
X = haralick_features(img, region=bw,distance=7)
elif fxi == 'fourier':
X = fourier_features(img, region=bw)
elif fxi == 'basicint':
X = basic_int_features(img, region=bw)
elif fxi == 'clp':
X = clp_features(img)
elif fxi == 'contrast':
X = contrast_features(img, region=bw)
elif fxi == 'dct':
X = dct_features(img, region=bw)
elif fxi == 'basicgeo':
X = basic_geo_features(bw)
elif fxi == 'centroid':
f = basic_geo_features(bw)
X = f[0:2]
elif fxi == 'fourierdes':
X = fourier_des_features(bw)
elif fxi == 'flusser':
X = flusser_features(bw)
elif fxi == 'gupta':
X = gupta_features(bw)
elif fxi == 'hugeo':
X = hugeo_features(bw)
elif fxi == 'ellipse':
X = ellipse_features(bw)
else:
print('ERROR: ' + fxi + ' does not exist as geometric feature extraction method.')
features = np.asarray(X)
return features
for filename in os.listdir(directory):
if filename.endswith(".png"):
# Para un archivo "face_xxx_nn.png", ID es xxx y nn es nn, ambos en int
id_nn = filename[:-4].split("_")
ID = int(id_nn[1])
nn = int(id_nn[2])
if ID % 2 == 1: # Nos quedamos con los impares
if nn <= 7: # y solo los 7 primeros
i += 1
# Leemos la imágen y obtenemos sus features dadas por lbp
im = imread(f"{directory}/{filename}")
lbp = lbp_features(im, hdiv=hdiv, vdiv=vdiv)
# Guardamos los resultados
LBPs.append(lbp)
IDs.append(ID)
NNs.append(nn)
if i % 10 == 0:
progressBar(i, 350, bar_length=20)
print(f"\nLBPs Calculado\nLBP Shape: {len(LBPs)}, {len(LBPs[0])}\nIDs" +
f"Shape: {len(IDs)}, 1")
# 2) Antes de la selección de features, realizaremos un cleansing con Clean.
# Luego realizaremos SFS.
Xtrain, Xtest, Ytrain, Ytest = [], [], [], []