def Varidation():
print("start predict")
size = (512, 512)
target_size = (size[0], size[1], 1)
filename = "image11_with_tumor"
path = "./image_data/"+filename
cimage, gimage, sizedata = load_image(path, size)
os.system("cls")
# with open("./image_data/predict_gray.pkl",mode='rb') as f:
# gimage = pickle.load(f)
# with open("./image_data/predict_clr.pkl",mode='rb') as f:
# cimage = pickle.load(f)
#model = unet_2d(input_shape=target_size)
model = unet_2d_GAM(input_shape=target_size)
model.load_weights('./model/unet_2d_weights.hdf5')
os.system("cls")
for i in range(len(cimage)):
mean = 0
print("\r{0:d}".format(i), end="")
w, h, c = gimage[i].shape
target = gimage[i].reshape((1, w, h, 1))
target = target / 255
pred = model.predict(target)[0]
pred = np.reshape(pred, (w, h, 1), order='F')
# 閾値
#mean = np.mean(pred)
mean = 0.95
# print(pred)
pred[pred >= mean] = 255
pred[pred < mean] = 0
pred = np.asarray(pred, dtype=np.uint8)
val = cimage[i]
"""
val :検出した領域の輪郭をoriginal画像と重ねた画像
label :検出した領域の輪郭
val :検出した領域
"""
val, label, _ = draw_contours(val, pred)
cv2.imwrite("./result/test/"+filename +
"/target/{0:04d}.jpg".format(i), cimage[i])
cv2.imwrite("./result/test/"+filename +
"/pred/{0:04d}.jpg".format(i), pred)
cv2.imwrite("./result/test/"+filename +
"/val/{0:04d}.jpg".format(i), val)
def test(dict):
"""
# # Amed Test Imageで抽出器の性能を評価する #
Parameter
---------
dict :辞書型
-Parameter.txtから必要な情報を格納したdict
Result
------
抽出結果を //aka/share/amed/amed_unet_result に書き出し
"""
print("start predict")
size = dict["size"]
root = dict["amedroot_path"]
target_size = (size[0], size[1], 1)
cimage, gimage, data_name = load_image_from_txt(dict)
#cimage,gimage = load_image_s('./image/image.pkl')
model = unet_2d_GAM(input_shape=target_size)
model.load_weights(dict["model_path"])
os.system("cls")
for i in range(len(cimage)):
#start = time.time()
print("\r{0:d}".format(i), end="") #
w, h, c = gimage[i].shape
target = cv2.resize(gimage[i], size)
target = target.reshape((1, size[0], size[1], 1))
target = target / 255 #正規化
pred = model.predict(target)[0]
pred = np.reshape(pred, (size[0], size[1], 1), order='F')
#閾値
Threshold = 0.7
#推測
pred[pred >= Threshold] = 1
pred[pred < Threshold] = 0
pred = np.asarray(pred, dtype=np.uint8)
val = np.copy(cimage[i])
"""
val :検出した領域の輪郭をoriginal画像と重ねた画像
label :検出した領域の輪郭
val :検出した領域
"""
pred_resize = cv2.resize(pred, (h, w))
pred_resize = np.reshape(pred_resize, (w, h, 1), order='F')
val, label, relabel = draw_contours(val, pred_resize)
#pred_resize_2 = np.concatenate((pred_resize,pred_resize),axis=2)
#pred_resize = np.concatenate((pred_resize_2,pred_resize),axis=2)
result_image = (relabel / 255) * cimage[i]
#1枚の処理時間
#elapsed_time = time.time() - start
#print ("elapsed_time:{0}".format(elapsed_time) + "[sec]")
#print("pred_resize.shape",pred_resize.shape)
#print("cimage[i].shape",cimage[i].shape)
#save result
result_rootpath = "/amed_unet_result/" + os.path.dirname(data_name[i])
result_dir_path = root + result_rootpath
check_existfile(result_dir_path)
check_existfile(root + result_rootpath + "/val")
check_existfile(root + result_rootpath + "/target")
check_existfile(root + result_rootpath + "/pred")
cv2.imwrite(
root + result_rootpath + "/val/val_" +
os.path.basename(data_name[i]), val)
cv2.imwrite(root + "/amed_unet_result/" + data_name[i],
result_image * 255)
cv2.imwrite(
root + result_rootpath + "/target/target_" +
os.path.basename(data_name[i]), cimage[i])
cv2.imwrite(
root + result_rootpath + "/pred/" + os.path.basename(data_name[i]),
relabel * 255)
x1,x2,y1,y2 = 200,500,200,500
show(tumor_coord)
#座標受け取り
#x1,x2,y1,y2 = coord
tumor_coord[y1:y2,x1:x2] = 1
result = tumor_coord * pred
tumor_area= np.count_nonzero(tumor_coord >= 1)#腫瘍領域の面積
result_area = np.count_nonzero(result >= 1)#重複する面積
Inclusion_rate = result_area / tumor_area
print(Inclusion_rate)
show(pred)
show(result)
input()
if __name__ == "__main__":
test_amed()
input()
target_size = (640,640,1)
#model
seg_model = unet_2d_GAM(input_shape=target_size)
seg_model.load_weights("./model/unet_2d_weights.hdf5")
Segmentation(image,seg_model,target_size)
def test_amed():
"""
"""
print("start predict")
size = (512, 512)
target_size = (size[0], size[1], 1)
filename = "image06"
path = "./image_data/"+filename
cimage, gimage, sizedata = load_image(path, size)
# with open("./image_data/predict_gray.pkl",mode='rb') as f:
# gimage = pickle.load(f)
# with open("./image_data/predict_clr.pkl",mode='rb') as f:
# cimage = pickle.load(f)
#model = unet_2d(input_shape=target_size)
model = unet_2d_GAM(input_shape=target_size)
model.load_weights('./model/unet_2d_weights.hdf5')
os.system("cls")
for i in range(len(cimage)):
mean = 0
start = time.time()
print("\r{0:d}".format(i), end="")
w, h, c = gimage[i].shape
target = cv2.resize(gimage[i], size)
target = target.reshape((1, 512, 512, 1))
target = target / 255
pred = model.predict(target)[0]
pred = np.reshape(pred, (512, 512, 1), order='F')
# Heat Map 作成
#pred_heat = pred[:,:,0]
gray = np.copy(gimage[i])
# 疑似カラーを付与
heatmap = np.round(pred*255)
heatmap = heatmap.astype('uint8')
heatmap = cv2.applyColorMap(heatmap, 2)
# 画像合成
heatmap = cv2.resize(heatmap, (h, w))
# チャネル追加
gray_2 = np.concatenate((gray, gray), axis=2)
gray = np.concatenate((gray_2, gray), axis=2)
heatmap = cv2.addWeighted(gray, 0.3, heatmap, 0.7, 0)
check_existfile("./result/test/"+filename+"/heatmap")
cv2.imwrite("./result/test/"+filename +
"/heatmap/{0:04d}.jpg".format(i), heatmap)
#mean = np.mean(pred)
mean = 0.7
# 推測
pred[pred >= mean] = 1
pred[pred < mean] = 0
pred = np.asarray(pred, dtype=np.uint8)
val = np.copy(cimage[i])
"""
val :検出した領域の輪郭をoriginal画像と重ねた画像
label :検出した領域の輪郭
val :検出した領域
"""
pred_resize = cv2.resize(pred, (h, w))
pred_resize = np.reshape(pred_resize, (w, h, 1), order='F')
val, label, _ = draw_contours(val, pred_resize)
val_resize = cv2.resize(val, (h, w))
pred_resize_2 = np.concatenate((pred_resize, pred_resize), axis=2)
pred_resize = np.concatenate((pred_resize_2, pred_resize), axis=2)
#result_image = pred_resize*cimage[i]
elapsed_time = time.time() - start
print("elapsed_time:{0}".format(elapsed_time) + "[sec]")
#cv2.imwrite("//aka/share/amed/amed_liver_label/test_result/target/{0:04d}.jpg".format(i) , cimage[i])
#cv2.imwrite("//aka/share/amed/amed_liver_label/test_result/pred/{0:04d}.jpg".format(i) , pred_resize*255)
#cv2.imwrite("//aka/share/amed/amed_liver_label/test_result/val/{0:04d}.jpg".format(i) , val_resize)
cv2.imwrite("./result/test/"+filename +
"/target/{0:04d}.jpg".format(i), cimage[i])
cv2.imwrite("./result/test/"+filename +
"/pred/{0:04d}.jpg".format(i), pred_resize*255)
cv2.imwrite("./result/test/"+filename +
"/val/{0:04d}.jpg".format(i), val_resize)
def train():
"""
# # ネットワークモデルの学習をするための関数
-Network model :U-net
-save weight only
##注意点
-学習時の画像サイズを必ず記録する事
"""
"""
画像読み込み
前処理(サイズ統一、終わってたらスルー)
学習(fit)
検証(predict)
精度と学習モデル保存
"""
# Setup Parameter
size = (512, 512)
target_size = (size[0], size[1], 1)
batch_size = 4
epoch = 50
classes = 2
gen_num = 5
seed = 10
# 実行時間の計測用（処理実行開始時間）
dt_now = datetime.datetime.now()
days = dt_now.year+dt_now.month + dt_now.day
# Load model
#model = unet_2d(input_shape = target_size)
model = unet_2d_GAM(input_shape=target_size)
model.summary()
# Setup dataset
image = load_image_s('./image/image.pkl')
label = load_image_s('./image/label.pkl')
# preprocessing -前処理-
# 前処理に使うのはやめた方がいいかもしれない
#image,label = Echo_Extraction(image,label)
# for ss in range(len(image)):
# cv2.imwrite("./image/check/image/{0:04d}.jpg".format(ss) , image[ss])
# cv2.imwrite("./image/check/label/{0:04d}.jpg".format(ss), label[ss])
# 正規化
image = image/255
label = label/255
train_image, val_image = train_test_split(
image, test_size=0.3, random_state=seed, shuffle=True)
train_label, val_label = train_test_split(
label, test_size=0.3, random_state=seed, shuffle=True)
# check dataset
"""
t_image = train_image * 255
t_label = train_label * 255
v_image = val_image * 255
v_label = val_label * 255
for i in range(len(train_image)):
cv2.imwrite("./image/check/train_image/{0:04d}.jpg".format(i) , t_image[i])
cv2.imwrite("./image/check/train_label/{0:04d}.jpg".format(i), t_label[i])
for i in range(len(val_image)):
cv2.imwrite("./image/check/val_image/{0:04d}.jpg".format(i), v_image[i])
cv2.imwrite("./image/check/val_label/{0:04d}.jpg".format(i), v_label[i])
"""
# Data Augumentation
nb_data = len(train_image) * gen_num
val_nb_data = len(val_image) * gen_num
#train_gen = ImageDataGenerator(preprocessing_function = myprocess_pixel_value,width_shift_range=[-10,10],height_shift_range=[-10,10])
train_gen = ImageDataGenerator(
width_shift_range=0.1, height_shift_range=0.1)
val_gen = ImageDataGenerator()
# Setup CallBacks
ModelCheckpoint_ = ModelCheckpoint(
'./model/unet_2d_weights.hdf5', monitor='val_loss', save_weights_only=True)
early_stopping = EarlyStopping(monitor='val_loss', patience=5, verbose=1)
ReduceLROnPlateau_ = ReduceLROnPlateau(
monitor='val_loss', factor=0.2, patience=3, verbose=1, mode='auto', epsilon=0.001, cooldown=0, min_lr=0.002)
CSVLogger_ = keras.callbacks.CSVLogger(
"./model/callback/log_learning.csv", separator=',', append=False)
TensorBoard_ = keras.callbacks.TensorBoard(
log_dir="./model/log", histogram_freq=1)
# Set CallBack
#callbacks = [ModelCheckpoint_,ReduceLROnPlateau_,early_stopping,CSVLogger_]
callbacks = [ModelCheckpoint_, early_stopping, CSVLogger_]
#callbacks = [ModelCheckpoint_]
# Learning
# (generator=train_gen.flow(train_image, train_label,batch_size=batch_size,save_to_dir="E:/output/data_augumentation/1", save_prefix='img', save_format='jpg')
model.fit_generator(generator=train_gen.flow(train_image, train_label, batch_size=batch_size),
steps_per_epoch=nb_data,
epochs=epoch,
validation_data=val_gen.flow(
val_image, val_label, batch_size=batch_size),
validation_steps=val_nb_data,
max_queue_size=10,
callbacks=callbacks)
# log保存
with open("./model/model_detail.txt", mode='w') as f_dataset:
f_dataset.write("作成日時 : " + days)
f_dataset.write("画像サイズ : " + size)
f_dataset.write("seed : " + seed)
f_dataset.write("使用モデル : unet_2d_GAM")
f_dataset.write("\n画像枚数 :" + str(len(image)))
f_dataset.write("\n - train_image : " + str(len(train_image)))
f_dataset.write("\n - val_image : " + str(len(val_image)))
# 以下検証
# 重みの保存位置修正
model.load_weights('./model/unet_2d_weights.hdf5')
score = model.evaluate(val_image, val_label, batch_size=3, verbose=1)
print('Test loss ', score[0])
print('Test accuracy ', score[1])