def __init__(self):
self.im_heigth = 8 * 40
self.im_width = 16 * 40
input_shape = (self.im_heigth, self.im_width, 1)
self.maskModel = getSegModel(input_shape)
self.ocrModel = getOCRModel()
self.ocrModel.load_weights(currPath+'/checkpoints/OCRmodel_vl0.7638.hdf5')
def locnetWithMask(input_shape):
segmodel = getSegModel(input_shape)
for layer in segmodel.layers:
layer.trainable = False
inp = Input(input_shape)
segmentMask = segmodel(inp)
combinedinput = concatenate([inp,segmentMask], axis=-1)
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(64, kernel_size=3, padding='same')(combinedinput))))
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(64, kernel_size=3, padding='same')(x))))
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(128, kernel_size=3, padding='same')(x))))
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(128, kernel_size=3, padding='same')(x))))
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(256, kernel_size=3, padding='same')(x))))
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(256, kernel_size=3, padding='same')(x))))
xZ = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(256, kernel_size=3, padding='same')(x))))
x = Flatten()(xZ)
x = Dense(256, name='emb1', activation='relu')(x)
x = Dropout(0.25)(x)
x = Dense(256, name='emb2', activation='relu')(x)
x = Dropout(0.25)(x)
x = Dense(6, name='out')(x)
locnet = Model(input=inp, output=x)
locnet.load_weights('checkpoints/LOCNET_vl0.0469.hdf5')
locnet.compile('adam', 'mse')
for layer in locnet.layers:
layer.trainable = False
locnet.summary()
return locnet
def locnet_and_sp_mask(input_shape):
segmodel = getSegModel(input_shape, compile=False)
transformedInputSize = (32, 16 * 32, 1)
# for layer in segmodel.layers:
# layer.trainable = False
inp = Input(input_shape)
segmentMask = segmodel(inp)
combinedinput = concatenate([inp, segmentMask], axis=-1)
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(32, kernel_size=3, padding='same')(segmentMask))))
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(32, kernel_size=3, padding='same')(x))))
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(64, kernel_size=3, padding='same')(x))))
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(64, kernel_size=3, padding='same')(x))))
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(128, kernel_size=3, padding='same')(x))))
x = MaxPooling2D((2, 2))(LeakyReLU(0.2)(BatchNormalization()(Conv2D(128, kernel_size=3, padding='same')(x))))
xZ = LeakyReLU(0.2)(BatchNormalization()(Conv2D(128, kernel_size=3, padding='same')(x)))
x = LeakyReLU(0.2)(BatchNormalization()(Conv2D(64, kernel_size=3, padding='same')(xZ)))
x = Reshape((2048,))(x)
x = Dropout(0.25)(x)
x = LeakyReLU(0.2)(Dense(256, name='emb2')(x))
x = Dropout(0.25)(x)
x = Dense(6, name='out')(x)
locnet = Model(input=inp, output=x)
#locnet.compile('adam','mse')
#locnet.load_weights('checkpoints/LOCNET_vl0.0469.hdf5')
transformedMask = SpatialTransformer(localization_net=locnet, output_size=transformedInputSize[:2], input_shape=input_shape)(segmentMask)
model = Model(input = inp, outputs=[x, transformedMask, segmentMask])
model.compile('sgd', ['mse', 'mse', 'binary_crossentropy'], loss_weights=[1.,0.,1.])
model.summary()
return model
def locnet_on_mask(shape, retunCrop=False):
transformedInputSize = (32, 16 * 32, 1)
locnet = getlocnet(shape)
inp = Input(shape)
transformedInput = SpatialTransformer(localization_net=locnet,
output_size=transformedInputSize[:2],
input_shape=shape)(inp)
segmodel = getSegModel(transformedInputSize)
for layer in segmodel.layers:
layer.trainable = False
segOutput = segmodel(transformedInput)
if retunCrop:
endToEndModel = Model(input=inp, output=[segOutput, transformedInput])
else:
endToEndModel = Model(input=inp, output=segOutput)
endToEndModel.compile('sgd', 'binary_crossentropy')
return endToEndModel
import numpy as np
from scipy.misc import imsave
from utils import crop_rotate, mirror
import cv2
import os
import math
import itertools
from scipy.misc import imsave
from adjustTextPosition import adjust
im_heigth = 8 * 40
im_width = 16 * 40
input_shape = (im_heigth, im_width, 1)
max_angle = 45
model = getSegModel(input_shape)
model.load_weights('checkpoints/segmenter_vl0.0163.hdf5')
imgs = [
os.path.join(root, f) for root, _, files in os.walk('./tmp') for f in files
if f.lower().endswith('.jpg')
]
for imgpath in imgs:
savepath = imgpath.replace('tmp', 'trainimgs')
if not os.path.exists(savepath):
img = Image.open(imgpath)
img = img.resize((im_width, im_heigth))
img = np.array(img.convert('L'))
np_img = img / 127.5 - 1