def cross_validation_segnet():
from sklearn.model_selection import KFold
from sklearn.model_selection import train_test_split
from src.plot import plot_loss_accuracy, plot_dice_coefficient_cv
from src.timer import Timer
segmentation_test = True
rotation = False
jaccards_previous = []
jaccards_proposed = []
# training_validation dataset path
training_validation_path = 'datasets' + os.sep + 'training_validation'
# 学習・検証用データセットのロード
# load dataset for training, validation
x_all, file_names = load_x(training_validation_path + os.sep + 'image')
y_all = load_y(training_validation_path + os.sep + 'label')
# testデータの分割だがこれは今回あらかじめ分けておくので無視する
#x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=7)
# 指領域抽出実験用
X_test, file_names = load_x(
'datasets' + os.sep + 'segmentation_test' + os.sep + 'image', rotation)
# 入力はグレースケール1チャンネル
input_channel_count = 1
# 出力はグレースケール1チャンネル
output_channel_count = 1
# hyper parameters
BATCH_SIZE = 8
NUM_EPOCH = 300
# dice係数の最終値を記憶するlist
# lists for get last dice coefficient values
final_dices = []
final_val_dices = []
# dice係数のlistを記憶するlist
# list for get dice coefficient
dice_lists = []
training_times = []
label_names = ['1st', '2nd', '3rd', '4th']
kf = KFold(n_splits=4, shuffle=True)
timer = Timer()
# kFoldループを行う(36データを4つに分割)
# kFold loop (split 36 data to 4 data)
for train_index, val_index in kf.split(x_all, y_all):
print(train_index, val_index)
x_train = x_all[train_index]
y_train = y_all[train_index]
x_validation = x_all[val_index]
y_validation = y_all[val_index]
print(x_train.shape, x_validation.shape)
# SegNetの定義
model = segnet(input_channel_count, output_channel_count,
FIRST_LAYER_FILTER_COUNT)
model.compile(loss=dice_coefficient_loss,
optimizer=Adam(lr=1e-3),
metrics=[dice_coefficient, 'accuracy'])
timer.reset()
history = model.fit(x_train,
y_train,
batch_size=BATCH_SIZE,
epochs=NUM_EPOCH,
verbose=1,
validation_data=(x_validation, y_validation))
training_times.append(timer.time_elapsed())
# dice_coefficientの最終値を記録
# get last dice coefficient value
final_dices.append(history.history['dice_coefficient'][-1])
final_val_dices.append(history.history['val_dice_coefficient'][-1])
dice_lists.append(history.history['dice_coefficient'])
plot_loss_accuracy(history)
#Y_pred = model.predict(X_test, BATCH_SIZE)
if segmentation_test:
Y_pred = model.predict(X_test, BATCH_SIZE)
for i, y in enumerate(Y_pred):
# testDataフォルダ配下にleft_imagesフォルダを置いている
img = cv2.imread(
'datasets' + os.sep + 'segmentation_test' + os.sep +
'image' + os.sep + file_names[i], 0)
if rotation:
y = cv2.resize(y, (img.shape[0], img.shape[0]))
else:
y = cv2.resize(y, (img.shape[1], img.shape[0]))
# 提案手法
# proposed method
y_dn = denormalize_y(y)
y_dn = np.uint8(y_dn)
# ret, mask = cv2.threshold(y_dn, 0, 255, cv2.THRESH_OTSU)
ret, mask = cv2.threshold(y_dn, 127, 255, cv2.THRESH_BINARY)
# hist, bins = np.histogram(mask.ravel(), 256, [0, 256])
mask_binary = normalize_y(mask)
# 従来手法
# previous method
mask_previous, _ = opening_masking(img)
mask_previous_binary = normalize_y(mask_previous)
# ラベル読み込み
# load label images
label = cv2.imread(
'datasets' + os.sep + 'segmentation_test' + os.sep +
'label' + os.sep + file_names[i], 0)
label_binary = normalize_y(label)
dice_previous = K.get_value(
jaccard_coefficient(mask_previous_binary, label_binary))
dice_proposed = K.get_value(
jaccard_coefficient(mask_binary, label_binary))
print('previous:', dice_previous)
jaccards_previous.append(dice_previous)
print('proposed:', dice_proposed)
jaccards_proposed.append(dice_proposed)
print(final_dices)
print('平均訓練精度', np.mean(final_dices)) # mean training precision
print(final_val_dices)
print('平均検証精度', np.mean(final_val_dices)) # mean validation precision
print(training_times)
print('平均学習時間', np.mean(training_times)) # mean training time
print(jaccards_previous)
print('従来手法精度',
np.mean(jaccards_previous)) # extract-precision by previous method
print(jaccards_proposed)
print('提案手法精度',
np.mean(jaccards_proposed)) # extract-precision by proposed method
plot_dice_coefficient_cv(dice_lists, label_names)
return 0
def segnet_predict():
import cv2
from src.timer import Timer
rotation = False
# set True for running segmentation test
segmentation_test = True
jaccard_sum_previous = 0
jaccard_sum_proposed = 0
time_previous = [0, 0, 0, 0]
time_proposed = 0
if segmentation_test:
# 指領域抽出実験用
# for segmentation test
X_test, file_names = load_x(
'datasets' + os.sep + 'segmentation_test' + os.sep + 'image',
rotation)
else:
# normal predict
X_test, file_names = load_x(
'datasets' + os.sep + 'test' + os.sep + 'image', rotation)
input_channel_count = 1
output_channel_count = 1
first_layer_filter_count = FIRST_LAYER_FILTER_COUNT
model = segnet(input_channel_count, output_channel_count,
first_layer_filter_count)
model.load_weights('segnet_weights.hdf5')
model.summary()
BATCH_SIZE = 8
timer = Timer()
Y_pred = model.predict(X_test, BATCH_SIZE)
#time_proposed += timer.time_elapsed()
for i, y in enumerate(Y_pred):
# testDataフォルダ配下にleft_imagesフォルダを置いている
img = cv2.imread(
'datasets' + os.sep + 'test' + os.sep + 'image' + os.sep +
file_names[i], 0)
#cv2.equalizeHist(img ,img)
if rotation:
y = cv2.resize(y, (img.shape[0], img.shape[0]))
else:
y = cv2.resize(y, (img.shape[1], img.shape[0]))
y_dn = denormalize_y(y)
y_dn = np.uint8(y_dn)
#ret, mask = cv2.threshold(y_dn, 0, 255, cv2.THRESH_OTSU)
ret, mask = cv2.threshold(y_dn, 127, 255, cv2.THRESH_BINARY)
#hist, bins = np.histogram(mask.ravel(), 256, [0, 256])
mask_binary = normalize_y(mask)
timer.reset()
masked = cv2.bitwise_and(img, mask)
mask_rest = cv2.bitwise_not(mask)
masked = cv2.bitwise_or(masked, mask_rest)
#image_user_processed = high_boost_filter(masked)
cv2.imwrite('prediction' + os.sep + file_names[i], mask)
time_proposed += timer.time_elapsed()
if segmentation_test:
img = cv2.imread(
'datasets' + os.sep + 'segmentation_test' + os.sep + 'image' +
os.sep + file_names[i], 0)
mask_previous, _, time_list_previous = opening_masking(img)
time_previous[0] += time_list_previous[0]
time_previous[1] += time_list_previous[1] - time_list_previous[0]
time_previous[2] += time_list_previous[2] - time_list_previous[1]
time_previous[3] += time_list_previous[3] - time_list_previous[2]
mask_previous_binary = normalize_y(mask_previous)
print(mask_previous_binary.shape)
label = cv2.imread(
'datasets' + os.sep + 'segmentation_test' + os.sep + 'label' +
os.sep + file_names[i], 0)
label_binary = normalize_y(label)
jaccard_previous = K.get_value(
jaccard_coefficient(mask_previous_binary, label_binary))
jaccard_proposed = K.get_value(
jaccard_coefficient(mask_binary, label_binary))
print('previous:', jaccard_previous)
jaccard_sum_previous += jaccard_previous
print('proposed:', jaccard_proposed)
jaccard_sum_proposed += jaccard_proposed
print('sum_previous:', jaccard_sum_previous)
print('sum_proposed:', jaccard_sum_proposed)
print('mean_previous:', jaccard_sum_previous / len(Y_pred))
print('mean_proposed:', jaccard_sum_proposed / len(Y_pred))
print('mean_time_previous:',
list(map(lambda x: x / len(Y_pred), time_previous)))
print('mean_time_proposed', time_proposed / len(Y_pred))
return 0
