使用 Tesseract OCR 和 python 进行数字识别
Digit recognition with Tesseract OCR and python
我使用 Tesseract 和 python 读取数字(从电表)。
除数字“1”外,一切正常。
Tesseract 无法读取数字“1”。
这是我发给 tesseract 的图片:
超正方体显示为“0000027”。
如何告诉 Tesseract 垂直杆是“1”?
这是我的 tesseract 初始化:
import tesseract
TESSERACT_LIBRARY_PATH = "C:\Program Files (x86)\Tesseract-OCR"
LANGUAGE = "eng"
CHARACTERS = "0123456789"
FALSE = "0"
TRUE = "1"
def init_ocr():
"""
.. py:function:: init_ocr()
Utilize the Tesseract-OCR library to create an tesseract_ocr that
predicts the numbers to be read off of the meter.
:return: tesseract_ocr Tesseracts OCR API.
:rtype: Class
"""
# Initialize the tesseract_ocr with the english language package.
tesseract_ocr = tesseract.TessBaseAPI()
tesseract_ocr.Init(TESSERACT_LIBRARY_PATH, LANGUAGE,
tesseract.OEM_DEFAULT)
# Limit the characters being seached for to numerics.
tesseract_ocr.SetVariable("tessedit_char_whitelist", CHARACTERS)
# Set the tesseract_ocr to predict for only one character.
tesseract_ocr.SetPageSegMode(tesseract.PSM_AUTO)
# Tesseract's Directed Acyclic Graph.
# Not necessary for number recognition.
tesseract_ocr.SetVariable("load_system_dawg", FALSE)
tesseract_ocr.SetVariable("load_freq_dawg", FALSE)
tesseract_ocr.SetVariable("load_number_dawg", TRUE)
tesseract_ocr.SetVariable("classify_enable_learning", FALSE)
tesseract_ocr.SetVariable("classify_enable_adaptive_matcher", FALSE)
return tesseract_ocr
略微无关紧要的答案,但可能有助于您的最初目标。
我在使用 tesseract 时遇到了类似的问题,而且我对性能的要求也非常严格。我在 SO 上找到了 this 简单的解决方案,并使用 OpenCV 制作了简单的识别器。
它归结为在您拥有的非常清晰的图像上找到边界矩形(从边缘),然后尝试将找到的对象与模板进行匹配。我相信您的解决方案既简单又精确,但需要的代码比现在多一些。
我会关注这个问题,因为使用 tesseract 有可行的解决方案会很好。
我的时间有限,但这似乎是一个可行的解决方案:
import os
import cv2
import numpy
KNN_SQUARE_SIDE = 50 # Square 50 x 50 px.
def resize(cv_image, factor):
new_size = tuple(map(lambda x: x * factor, cv_image.shape[::-1]))
return cv2.resize(cv_image, new_size)
def crop(cv_image, box):
x0, y0, x1, y1 = box
return cv_image[y0:y1, x0:x1]
def draw_box(cv_image, box):
x0, y0, x1, y1 = box
cv2.rectangle(cv_image, (x0, y0), (x1, y1), (0, 0, 255), 2)
def draw_boxes_and_show(cv_image, boxes, title='N'):
temp_image = cv2.cvtColor(cv_image, cv2.COLOR_GRAY2RGB)
for box in boxes:
draw_box(temp_image, box)
cv2.imshow(title, temp_image)
cv2.waitKey(0)
class BaseKnnMatcher(object):
distance_threshold = 0
def __init__(self, source_dir):
self.model, self.label_map = self.get_model_and_label_map(source_dir)
@staticmethod
def get_model_and_label_map(source_dir):
responses = []
label_map = []
samples = numpy.empty((0, KNN_SQUARE_SIDE * KNN_SQUARE_SIDE), numpy.float32)
for label_idx, filename in enumerate(os.listdir(source_dir)):
label = filename[:filename.index('.png')]
label_map.append(label)
responses.append(label_idx)
image = cv2.imread(os.path.join(source_dir, filename), 0)
suit_image_standard_size = cv2.resize(image, (KNN_SQUARE_SIDE, KNN_SQUARE_SIDE))
sample = suit_image_standard_size.reshape((1, KNN_SQUARE_SIDE * KNN_SQUARE_SIDE))
samples = numpy.append(samples, sample, 0)
responses = numpy.array(responses, numpy.float32)
responses = responses.reshape((responses.size, 1))
model = cv2.KNearest()
model.train(samples, responses)
return model, label_map
def predict(self, image):
image_standard_size = cv2.resize(image, (KNN_SQUARE_SIDE, KNN_SQUARE_SIDE))
image_standard_size = numpy.float32(image_standard_size.reshape((1, KNN_SQUARE_SIDE * KNN_SQUARE_SIDE)))
closest_class, results, neigh_resp, distance = self.model.find_nearest(image_standard_size, k=1)
if distance[0][0] > self.distance_threshold:
return None
return self.label_map[int(closest_class)]
class DigitKnnMatcher(BaseKnnMatcher):
distance_threshold = 10 ** 10
class MeterValueReader(object):
def __init__(self):
self.digit_knn_matcher = DigitKnnMatcher(source_dir='templates')
@classmethod
def get_symbol_boxes(cls, cv_image):
ret, thresh = cv2.threshold(cv_image.copy(), 150, 255, cv2.THRESH_BINARY)
contours, hierarchy = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
symbol_boxes = []
for contour in contours:
x, y, width, height = cv2.boundingRect(contour)
# You can test here for box size, though not required in your example:
# if cls.is_size_of_digit(width, height):
# symbol_boxes.append((x, y, x+width, y+height))
symbol_boxes.append((x, y, x+width, y+height))
return symbol_boxes
def get_value(self, meter_cv2_image):
symbol_boxes = self.get_symbol_boxes(meter_cv2_image)
symbol_boxes.sort() # x is first in tuple
symbols = []
for box in symbol_boxes:
symbol = self.digit_knn_matcher.predict(crop(meter_cv2_image, box))
symbols.append(symbol)
return symbols
if __name__ == '__main__':
# If you want to see how boxes detection works, uncomment these:
# img_bw = cv2.imread(os.path.join('original.png'), 0)
# boxes = MeterValueReader.get_symbol_boxes(img_bw)
# draw_boxes_and_show(img_bw, boxes)
# Uncomment to generate templates from image
# import random
# TEMPLATE_DIR = 'templates'
# img_bw = cv2.imread(os.path.join('original.png'), 0)
# boxes = MeterValueReader.get_symbol_boxes(img_bw)
# for box in boxes:
# # You need to label templates manually after extraction
# cv2.imwrite(os.path.join(TEMPLATE_DIR, '%s.png' % random.randint(0, 1000)), crop(img_bw, box))
img_bw = cv2.imread(os.path.join('original.png'), 0)
vr = MeterValueReader()
print vr.get_value(img_bw)
我使用 Tesseract 和 python 读取数字(从电表)。 除数字“1”外,一切正常。 Tesseract 无法读取数字“1”。
这是我发给 tesseract 的图片:
超正方体显示为“0000027”。
如何告诉 Tesseract 垂直杆是“1”?
这是我的 tesseract 初始化:
import tesseract
TESSERACT_LIBRARY_PATH = "C:\Program Files (x86)\Tesseract-OCR"
LANGUAGE = "eng"
CHARACTERS = "0123456789"
FALSE = "0"
TRUE = "1"
def init_ocr():
"""
.. py:function:: init_ocr()
Utilize the Tesseract-OCR library to create an tesseract_ocr that
predicts the numbers to be read off of the meter.
:return: tesseract_ocr Tesseracts OCR API.
:rtype: Class
"""
# Initialize the tesseract_ocr with the english language package.
tesseract_ocr = tesseract.TessBaseAPI()
tesseract_ocr.Init(TESSERACT_LIBRARY_PATH, LANGUAGE,
tesseract.OEM_DEFAULT)
# Limit the characters being seached for to numerics.
tesseract_ocr.SetVariable("tessedit_char_whitelist", CHARACTERS)
# Set the tesseract_ocr to predict for only one character.
tesseract_ocr.SetPageSegMode(tesseract.PSM_AUTO)
# Tesseract's Directed Acyclic Graph.
# Not necessary for number recognition.
tesseract_ocr.SetVariable("load_system_dawg", FALSE)
tesseract_ocr.SetVariable("load_freq_dawg", FALSE)
tesseract_ocr.SetVariable("load_number_dawg", TRUE)
tesseract_ocr.SetVariable("classify_enable_learning", FALSE)
tesseract_ocr.SetVariable("classify_enable_adaptive_matcher", FALSE)
return tesseract_ocr
略微无关紧要的答案,但可能有助于您的最初目标。
我在使用 tesseract 时遇到了类似的问题,而且我对性能的要求也非常严格。我在 SO 上找到了 this 简单的解决方案,并使用 OpenCV 制作了简单的识别器。
它归结为在您拥有的非常清晰的图像上找到边界矩形(从边缘),然后尝试将找到的对象与模板进行匹配。我相信您的解决方案既简单又精确,但需要的代码比现在多一些。
我会关注这个问题,因为使用 tesseract 有可行的解决方案会很好。
我的时间有限,但这似乎是一个可行的解决方案:
import os
import cv2
import numpy
KNN_SQUARE_SIDE = 50 # Square 50 x 50 px.
def resize(cv_image, factor):
new_size = tuple(map(lambda x: x * factor, cv_image.shape[::-1]))
return cv2.resize(cv_image, new_size)
def crop(cv_image, box):
x0, y0, x1, y1 = box
return cv_image[y0:y1, x0:x1]
def draw_box(cv_image, box):
x0, y0, x1, y1 = box
cv2.rectangle(cv_image, (x0, y0), (x1, y1), (0, 0, 255), 2)
def draw_boxes_and_show(cv_image, boxes, title='N'):
temp_image = cv2.cvtColor(cv_image, cv2.COLOR_GRAY2RGB)
for box in boxes:
draw_box(temp_image, box)
cv2.imshow(title, temp_image)
cv2.waitKey(0)
class BaseKnnMatcher(object):
distance_threshold = 0
def __init__(self, source_dir):
self.model, self.label_map = self.get_model_and_label_map(source_dir)
@staticmethod
def get_model_and_label_map(source_dir):
responses = []
label_map = []
samples = numpy.empty((0, KNN_SQUARE_SIDE * KNN_SQUARE_SIDE), numpy.float32)
for label_idx, filename in enumerate(os.listdir(source_dir)):
label = filename[:filename.index('.png')]
label_map.append(label)
responses.append(label_idx)
image = cv2.imread(os.path.join(source_dir, filename), 0)
suit_image_standard_size = cv2.resize(image, (KNN_SQUARE_SIDE, KNN_SQUARE_SIDE))
sample = suit_image_standard_size.reshape((1, KNN_SQUARE_SIDE * KNN_SQUARE_SIDE))
samples = numpy.append(samples, sample, 0)
responses = numpy.array(responses, numpy.float32)
responses = responses.reshape((responses.size, 1))
model = cv2.KNearest()
model.train(samples, responses)
return model, label_map
def predict(self, image):
image_standard_size = cv2.resize(image, (KNN_SQUARE_SIDE, KNN_SQUARE_SIDE))
image_standard_size = numpy.float32(image_standard_size.reshape((1, KNN_SQUARE_SIDE * KNN_SQUARE_SIDE)))
closest_class, results, neigh_resp, distance = self.model.find_nearest(image_standard_size, k=1)
if distance[0][0] > self.distance_threshold:
return None
return self.label_map[int(closest_class)]
class DigitKnnMatcher(BaseKnnMatcher):
distance_threshold = 10 ** 10
class MeterValueReader(object):
def __init__(self):
self.digit_knn_matcher = DigitKnnMatcher(source_dir='templates')
@classmethod
def get_symbol_boxes(cls, cv_image):
ret, thresh = cv2.threshold(cv_image.copy(), 150, 255, cv2.THRESH_BINARY)
contours, hierarchy = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
symbol_boxes = []
for contour in contours:
x, y, width, height = cv2.boundingRect(contour)
# You can test here for box size, though not required in your example:
# if cls.is_size_of_digit(width, height):
# symbol_boxes.append((x, y, x+width, y+height))
symbol_boxes.append((x, y, x+width, y+height))
return symbol_boxes
def get_value(self, meter_cv2_image):
symbol_boxes = self.get_symbol_boxes(meter_cv2_image)
symbol_boxes.sort() # x is first in tuple
symbols = []
for box in symbol_boxes:
symbol = self.digit_knn_matcher.predict(crop(meter_cv2_image, box))
symbols.append(symbol)
return symbols
if __name__ == '__main__':
# If you want to see how boxes detection works, uncomment these:
# img_bw = cv2.imread(os.path.join('original.png'), 0)
# boxes = MeterValueReader.get_symbol_boxes(img_bw)
# draw_boxes_and_show(img_bw, boxes)
# Uncomment to generate templates from image
# import random
# TEMPLATE_DIR = 'templates'
# img_bw = cv2.imread(os.path.join('original.png'), 0)
# boxes = MeterValueReader.get_symbol_boxes(img_bw)
# for box in boxes:
# # You need to label templates manually after extraction
# cv2.imwrite(os.path.join(TEMPLATE_DIR, '%s.png' % random.randint(0, 1000)), crop(img_bw, box))
img_bw = cv2.imread(os.path.join('original.png'), 0)
vr = MeterValueReader()
print vr.get_value(img_bw)