def GetForegroundImgContourInfo(FilePath):
'''
딥러닝 모델을 이용하여 배경을 제거한다.
전경의 외곽선 정보를 리턴한다.
:param FilePath:
:return:
'''
# 모델 입력 이미지 사이즈는 32의 배수가 되어야 한다.
ModelInputHeight = 640
ModelInputWidth = 640
Model_Path = "Models\\Background\\background.hdf4"
model = resnet50_unet(n_classes=2, input_height=ModelInputHeight, input_width=ModelInputWidth)
model.load_weights(Model_Path)
out = model.predict_segmentation(
inp=FilePath,
out_fname="out.png"
)
org = cv.imread(FilePath)
# 모델 아웃풋 사이즈는 입력 사이즈의 절반이다.
org = cv.resize(org, (320, 320))
out2 = np.array(out * 1, dtype=np.uint8)
# threshed = cv.adaptiveThreshold(out2, 255, cv.ADAPTIVE_THRESH_MEAN_C, cv.THRESH_BINARY, 3, 0)
out2 = cv.cvtColor(out2, cv.COLOR_GRAY2BGR)
#print(org.shape)
#print(out2.shape)
# cv.imshow("FilePath", org)
# cv.waitKey(0)
masked = org * out2
#cv.imshow("masked", masked)
#cv.waitKey(0)
maskedImg = cv.resize(masked, (Configuration.ImageSizeForPreprocess, Configuration.ImageSizeForPreprocess))
# 알약 외곽선을 추출한다.
highThreshold = 100
contours, lastIdx = GetPillContour(img_color =maskedImg, highThreshold = highThreshold, showFlag = False)
if contours is None:
print("contour is not found!!")
return contours, lastIdx
def train(images, labels, model, width, height, name):
if model == 'pspnet_101':
segmentation_model = pspnet_101(n_classes=16,
input_width=width,
input_height=height)
elif model == 'resnet50_unet':
segmentation_model = resnet50_unet(n_classes=16,
input_width=width,
input_height=height)
elif model == 'resnet50_segnet':
segmentation_model = resnet50_segnet(n_classes=16,
input_width=width,
input_height=height)
elif model == 'resnet50_pspnet':
segmentation_model = resnet50_pspnet(n_classes=16,
input_width=width,
input_height=height)
else:
return None
print('Start training')
checkpoint_directory = f'./checkpoints/{model}_{width}x{height}{name}'
if not os.path.exists(checkpoint_directory):
os.makedirs(checkpoint_directory)
segmentation_model.train(
train_images=images,
train_annotations=labels,
checkpoints_path=
f'{checkpoint_directory}/{model}_{width}x{height}{name}',
epochs=10,
batch_size=1)
return segmentation_model
original_image = "dataset/semantic_drone_dataset/original_images/001.jpg"
label_image_semantic = "dataset/semantic_drone_dataset/label_images_semantic/001.png"
fig, axs = plt.subplots(1,2, figsize=(16, 8), constrained_layout=True)
axs[0].imshow( Image.open(original_image))
# axs.grid(False)
label_image_semantic = Image.open(label_image_semantic)
axs[1].imshow( Image.open(original_image))
# label_image_semantic = np.asarray(label_image_semantic)
axs[1].imshow(label_image_semantic,alpha = 0.6)
# axs[1].grid(False)
n_classes = 23 # Aerial Semantic Segmentation Drone Dataset tree, gras, other vegetation, dirt, gravel, rocks, water, paved area, pool, person, dog, car, bicycle, roof, wall, fence, fence-pole, window, door, obstacle
model = resnet50_unet(n_classes=n_classes , input_height=416, input_width=608 )
epochs = 10
model.train(
train_images = "dataset/semantic_drone_dataset/original_images/",
train_annotations = "dataset/semantic_drone_dataset/label_images_semantic/",
checkpoints_path = "resnet50_unet" , epochs=epochs)
import time
from PIL import Image
import matplotlib.pyplot as plt
%matplotlib inline
start = time.time()
input_image = "dataset/semantic_drone_dataset/original_images/001.jpg"
def process(self):
try:
page_grp, self.image_grp = self.output_file_grp.split(',')
except ValueError:
page_grp = self.output_file_grp
self.image_grp = FALLBACK_IMAGE_GRP
LOG.info(
"No output file group for images specified, falling back to '%s'",
FALLBACK_IMAGE_GRP)
oplevel = self.parameter['operation_level']
model = None
if self.parameter['use_deeplr']:
model_weights = self.parameter['seg_weights']
if not Path(model_weights).is_file():
LOG.error("""\
Segementation model weights file was not found at '%s'. Make sure the `seg_weights` parameter
points to the local model weights path.
""" % model_weights)
sys.exit(1)
model = resnet50_unet(n_classes=self.parameter['classes'],
input_height=self.parameter['height'],
input_width=self.parameter['width'])
model.load_weights(model_weights)
LOG.info('Segmentation Model loaded')
for (n, input_file) in enumerate(self.input_files):
page_id = input_file.pageId or input_file.ID
pcgts = page_from_file(self.workspace.download_file(input_file))
metadata = pcgts.get_Metadata()
metadata.add_MetadataItem(
MetadataItemType(
type_="processingStep",
name=self.ocrd_tool['steps'][0],
value=TOOL,
Labels=[
LabelsType( #externalRef="parameters",
Label=[
LabelType(type_=name,
value=self.parameter[name])
for name in self.parameter.keys()
])
]))
page = pcgts.get_Page()
LOG.info("INPUT FILE %s", input_file.pageId or input_file.ID)
if self.parameter['use_deeplr']:
page_image, page_xywh, page_image_info = self.workspace.image_from_page(
page, page_id, feature_filter='binarized,deskewed,cropped')
else:
page_image, page_xywh, page_image_info = self.workspace.image_from_page(
page,
page_id,
feature_selector='binarized,deskewed,cropped')
if oplevel == 'page':
self._process_segment(page_image, page, page_xywh, page_id,
input_file, n, model)
else:
LOG.warning('Operation level %s, but should be "page".',
oplevel)
break
# Use input_file's basename for the new file -
# this way the files retain the same basenames:
file_id = input_file.ID.replace(self.input_file_grp,
self.output_file_grp)
if file_id == input_file.ID:
file_id = concat_padded(self.output_file_grp, n)
self.workspace.add_file(
ID=file_id,
file_grp=page_grp, #self.output_file_grp,
pageId=input_file.pageId,
mimetype=MIMETYPE_PAGE,
local_filename=os.path.join(self.output_file_grp,
file_id + '.xml'),
content=to_xml(pcgts).encode('utf-8'),
force=self.parameter['force'])
# 결과 이미지가 입력의 반으로 줄어듦(이유는 모름)
# 따라서 모델에다가 입력 크기를 원본보다 두배로 설정하면 결과가 원본 이미지와 같은 크기가 나옴.
# 크기는 32의 배수가 되어야 함. assert input_height%32 == 0
# 학습시 색 정보가 중요 한 것 같음.
# 배경과 전경의 색이 비슷하면 결과가 좋지 않음.
#from keras_segmentation.models.unet import vgg_unet
from keras_segmentation.models.unet import resnet50_unet
#model = vgg_unet(n_classes=2 , input_height=640, input_width=640)
model = resnet50_unet(n_classes=2, input_height=640, input_width=640)
model.train(
train_images="E:\\Study\\Pill\\SegTest\\Data\\firstDis2\\Training\\Src\\",
train_annotations=
"E:\\Study\\Pill\\SegTest\\Data\\firstDis2\\Training\\Mask\\",
checkpoints_path=
"E:\\Study\\Pill\\SegTest\\Data\\firstDis2\\Model\\Resnet50",
epochs=5)
import os, io, base64
from keras_segmentation.models.unet import resnet50_unet
from flask import Flask, request, jsonify
import numpy as np
import cv2
application = Flask(__name__)
model = resnet50_unet(n_classes=2, input_height=256, input_width=512)
model.load_weights("58.h5")
@application.route('/localOcr', methods=['POST'])
def recognition():
imgBase64 = str(request.get_data()).replace("b'image=", "") #直接数据的原始结构
imgBase64 = base64.b64decode(imgBase64)
imgBase64 = np.frombuffer(imgBase64, np.uint8)
imgBase64 = cv2.imdecode(imgBase64, cv2.IMREAD_ANYCOLOR)
out = model.predict_segmentation(inp=imgBase64, out_fname="3.png")
return jsonify(out.tolist())
if __name__ == "__main__":
application.run(host='0.0.0.0', port=80, debug=True, use_reloader=False)