DL之CNN：利用CNN(keras, CTC loss)算法实现OCR光学字符识别
目录
输出结果
实现的全部代码
输出结果
更新……
实现的全部代码
部分代码源自：GitHubhttps://raw.githubusercontent.com/fchollet/keras/master/examples/image_ocr.py
# -*- coding: utf-8 -*-
#image_ocr.py
'''
# Optical character recognition
This example uses a convolutional stack followed by a recurrent stack
and a CTC logloss function to perform optical character recognition
of generated text images. I have no evidence of whether it actually
learns general shapes of text, or just is able to recognize all
the different fonts thrown at it...the purpose is more to demonstrate CTC
inside of Keras. Note that the font list may need to be updated
for the particular OS in use.
This starts off with 4 letter words. For the first 12 epochs, the
difficulty is gradually increased using the TextImageGenerator class
which is both a generator class for test/train data and a Keras
callback class. After 20 epochs, longer sequences are thrown at it
by recompiling the model to handle a wider image and rebuilding
the word list to include two words separated by a space.
The table below shows normalized edit distance values. Theano uses
a slightly different CTC implementation, hence the different results.
Epoch | TF | TH
-----:|-------:|-------:
10| 0.027 | 0.064
15| 0.038 | 0.035
20| 0.043 | 0.045
25| 0.014 | 0.019
This requires ```cairo``` and ```editdistance``` packages:
```python
pip install cairocffi
pip install editdistance
```
Created by Mike Henry
https://github.com/mbhenry/
'''
import os
import itertools
import codecs
import re
import datetime
import cairocffi as cairo
import editdistance
import numpy as np
from scipy import ndimage
import pylab
from keras import backend as K
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras.layers import Input, Dense, Activation
from keras.layers import Reshape, Lambda
from keras.layers.merge import add, concatenate
from keras.models import Model
from keras.layers.recurrent import GRU
from keras.optimizers import SGD
from keras.utils.data_utils import get_file
from keras.preprocessing import image
import keras.callbacks
OUTPUT_DIR = 'image_ocr'
# character classes and matching regex filter
regex = r'^[a-z ]+$'
alphabet = u'abcdefghijklmnopqrstuvwxyz '
np.random.seed(55)
# this creates larger "blotches" of noise which look
# more realistic than just adding gaussian noise
# assumes greyscale with pixels ranging from 0 to 1
def speckle(img):
severity = np.random.uniform(0, 0.6)
blur = ndimage.gaussian_filter(np.random.randn(*img.shape) * severity, 1)
img_speck = (img + blur)
img_speck[img_speck > 1] = 1
img_speck[img_speck (w - 2 * border_w_h[1]) or box[3] > (h - 2 * border_w_h[0]):
raise IOError(('Could not fit string into image.'
'Max char count is too large for given image width.'))
# teach the RNN translational invariance by
# fitting text box randomly on canvas, with some room to rotate
max_shift_x = w - box[2] - border_w_h[0]
max_shift_y = h - box[3] - border_w_h[1]
top_left_x = np.random.randint(0, int(max_shift_x))
if ud:
top_left_y = np.random.randint(0, int(max_shift_y))
else:
top_left_y = h // 2
context.move_to(top_left_x - int(box[0]), top_left_y - int(box[1]))
context.set_source_rgb(0, 0, 0)
context.show_text(text)
buf = surface.get_data()
a = np.frombuffer(buf, np.uint8)
a.shape = (h, w, 4)
a = a[:, :, 0] # grab single channel
a = a.astype(np.float32) / 255
a = np.expand_dims(a, 0)
if rotate:
a = image.random_rotation(a, 3 * (w - top_left_x) / w + 1)
a = speckle(a)
return a
def shuffle_mats_or_lists(matrix_list, stop_ind=None):
ret = []
assert all([len(i) == len(matrix_list[0]) for i in matrix_list])
len_val = len(matrix_list[0])
if stop_ind is None:
stop_ind = len_val
assert stop_ind = self.num_words:
self.cur_val_index = self.val_split + self.cur_val_index % 32
yield ret
def on_train_begin(self, logs={}):
self.build_word_list(16000, 4, 1)
self.paint_func = lambda text: paint_text(
text, self.img_w, self.img_h,
rotate=False, ud=False, multi_fonts=False)
def on_epoch_begin(self, epoch, logs={}):
# rebind the paint function to implement curriculum learning
if 3 = 21 and self.max_string_len < 12:
self.build_word_list(32000, 12, 0.5)
# the actual loss calc occurs here despite it not being
# an internal Keras loss function
def ctc_lambda_func(args):
y_pred, labels, input_length, label_length = args
# the 2 is critical here since the first couple outputs of the RNN
# tend to be garbage:
y_pred = y_pred[:, 2:, :]
return K.ctc_batch_cost(labels, y_pred, input_length, label_length)
# For a real OCR application, this should be beam search with a dictionary
# and language model. For this example, best path is sufficient.
def decode_batch(test_func, word_batch):
out = test_func([word_batch])[0]
ret = []
for j in range(out.shape[0]):
out_best = list(np.argmax(out[j, 2:], 1))
out_best = [k for k, g in itertools.groupby(out_best)]
outstr = labels_to_text(out_best)
ret.append(outstr)
return ret
class VizCallback(keras.callbacks.Callback):
def __init__(self, run_name, test_func, text_img_gen, num_display_words=6):
self.test_func = test_func
self.output_dir = os.path.join(
OUTPUT_DIR, run_name)
self.text_img_gen = text_img_gen
self.num_display_words = num_display_words
if not os.path.exists(self.output_dir):
os.makedirs(self.output_dir)
def show_edit_distance(self, num):
num_left = num
mean_norm_ed = 0.0
mean_ed = 0.0
while num_left > 0:
word_batch = next(self.text_img_gen)[0]
num_proc = min(word_batch['the_input'].shape[0], num_left)
decoded_res = decode_batch(self.test_func,
word_batch['the_input'][0:num_proc])
for j in range(num_proc):
edit_dist = editdistance.eval(decoded_res[j],
word_batch['source_str'][j])
mean_ed += float(edit_dist)
mean_norm_ed += float(edit_dist) / len(word_batch['source_str'][j])
num_left -= num_proc
mean_norm_ed = mean_norm_ed / num
mean_ed = mean_ed / num
print('\nOut of %d samples: Mean edit distance:'
'%.3f Mean normalized edit distance: %0.3f'
% (num, mean_ed, mean_norm_ed))
def on_epoch_end(self, epoch, logs={}):
self.model.save_weights(
os.path.join(self.output_dir, 'weights%02d.h5' % (epoch)))
self.show_edit_distance(256)
word_batch = next(self.text_img_gen)[0]
res = decode_batch(self.test_func,
word_batch['the_input'][0:self.num_display_words])
if word_batch['the_input'][0].shape[0] < 256:
cols = 2
else:
cols = 1
for i in range(self.num_display_words):
pylab.subplot(self.num_display_words // cols, cols, i + 1)
if K.image_data_format() == 'channels_first':
the_input = word_batch['the_input'][i, 0, :, :]
else:
the_input = word_batch['the_input'][i, :, :, 0]
pylab.imshow(the_input.T, cmap='Greys_r')
pylab.xlabel(
'Truth = \'%s\'\nDecoded = \'%s\'' %
(word_batch['source_str'][i], res[i]))
fig = pylab.gcf()
fig.set_size_inches(10, 13)
pylab.savefig(os.path.join(self.output_dir, 'e%02d.png' % (epoch)))
pylab.close()
def train(run_name, start_epoch, stop_epoch, img_w):
# Input Parameters
img_h = 64
words_per_epoch = 16000
val_split = 0.2
val_words = int(words_per_epoch * (val_split))
# Network parameters
conv_filters = 16
kernel_size = (3, 3)
pool_size = 2
time_dense_size = 32
rnn_size = 512
minibatch_size = 32
if K.image_data_format() == 'channels_first':
input_shape = (1, img_w, img_h)
else:
input_shape = (img_w, img_h, 1)
fdir = os.path.dirname(
get_file('wordlists.tgz',
origin='http://www.mythic-ai.com/datasets/wordlists.tgz',
untar=True))
img_gen = TextImageGenerator(
monogram_file=os.path.join(fdir, 'wordlist_mono_clean.txt'),
bigram_file=os.path.join(fdir, 'wordlist_bi_clean.txt'),
minibatch_size=minibatch_size,
img_w=img_w,
img_h=img_h,
downsample_factor=(pool_size ** 2),
val_split=words_per_epoch - val_words)
act = 'relu'
input_data = Input(name='the_input', shape=input_shape, dtype='float32')
inner = Conv2D(conv_filters, kernel_size, padding='same',
activation=act, kernel_initializer='he_normal',
name='conv1')(input_data)
inner = MaxPooling2D(pool_size=(pool_size, pool_size), name='max1')(inner)
inner = Conv2D(conv_filters, kernel_size, padding='same',
activation=act, kernel_initializer='he_normal',
name='conv2')(inner)
inner = MaxPooling2D(pool_size=(pool_size, pool_size), name='max2')(inner)
conv_to_rnn_dims = (img_w // (pool_size ** 2),
(img_h // (pool_size ** 2)) * conv_filters)
inner = Reshape(target_shape=conv_to_rnn_dims, name='reshape')(inner)
# cuts down input size going into RNN:
inner = Dense(time_dense_size, activation=act, name='dense1')(inner)
# Two layers of bidirectional GRUs
# GRU seems to work as well, if not better than LSTM:
gru_1 = GRU(rnn_size, return_sequences=True,
kernel_initializer='he_normal', name='gru1')(inner)
gru_1b = GRU(rnn_size, return_sequences=True,
go_backwards=True, kernel_initializer='he_normal',
name='gru1_b')(inner)
gru1_merged = add([gru_1, gru_1b])
gru_2 = GRU(rnn_size, return_sequences=True,
kernel_initializer='he_normal', name='gru2')(gru1_merged)
gru_2b = GRU(rnn_size, return_sequences=True, go_backwards=True,
kernel_initializer='he_normal', name='gru2_b')(gru1_merged)
# transforms RNN output to character activations:
inner = Dense(img_gen.get_output_size(), kernel_initializer='he_normal',
name='dense2')(concatenate([gru_2, gru_2b]))
y_pred = Activation('softmax', name='softmax')(inner)
Model(inputs=input_data, outputs=y_pred).summary()
labels = Input(name='the_labels',
shape=[img_gen.absolute_max_string_len], dtype='float32')
input_length = Input(name='input_length', shape=[1], dtype='int64')
label_length = Input(name='label_length', shape=[1], dtype='int64')
# Keras doesn't currently support loss funcs with extra parameters
# so CTC loss is implemented in a lambda layer
loss_out = Lambda(
ctc_lambda_func, output_shape=(1,),
name='ctc')([y_pred, labels, input_length, label_length])
# clipnorm seems to speeds up convergence
sgd = SGD(lr=0.02, decay=1e-6, momentum=0.9, nesterov=True, clipnorm=5)
model = Model(inputs=[input_data, labels, input_length, label_length],
outputs=loss_out)
# the loss calc occurs elsewhere, so use a dummy lambda func for the loss
model.compile(loss={'ctc': lambda y_true, y_pred: y_pred}, optimizer=sgd)
if start_epoch > 0:
weight_file = os.path.join(
OUTPUT_DIR,
os.path.join(run_name, 'weights%02d.h5' % (start_epoch - 1)))
model.load_weights(weight_file)
# captures output of softmax so we can decode the output during visualization
test_func = K.function([input_data], [y_pred])
viz_cb = VizCallback(run_name, test_func, img_gen.next_val())
model.fit_generator(
generator=img_gen.next_train(),
steps_per_epoch=(words_per_epoch - val_words) // minibatch_size,
epochs=stop_epoch,
validation_data=img_gen.next_val(),
validation_steps=val_words // minibatch_size,
callbacks=[viz_cb, img_gen],
initial_epoch=start_epoch)
if __name__ == '__main__':
run_name = datetime.datetime.now().strftime('%Y:%m:%d:%H:%M:%S')
train(run_name, 0, 20, 128)
# increase to wider images and start at epoch 20.
# The learned weights are reloaded
train(run_name, 20, 25, 512)
