Tesseract文字训练,以及样本生成

前面用做Tesseract做文字识别的时候，一般网上教程称使用jTessBoxEditor训练（最终我试验发现对于中文的图片文字识别而言训练基本没什么卵用）

当然使用jTessBoxEditor训练新的文字还是可以的，当时我发现jTessBoxEditor训练基本的配置文件基本是文字以及文字的坐标于是我使用python脚本生成了对应的训练图片以及配置文件。先上个图：

yahei_font2.box配置文件box的内容如下：



生成的配置文件如下：





下面贴出python代码，大致原理大家看下代码跑一下一个没什么问题：

# -*- coding: utf-8 -*-
from PIL import Image
from PIL import ImageFont
from PIL import ImageDraw
import os

CREATE_PATH = 'F:/img_test/create_train_image/'
WIDTH = 700
HEIGHT = 900
# 正常字体的大小
# FONT_SIZE = 40
# FONT_SIZE = 16
# FONT_SIZE = 20
FONT_SIZE = 25
FONT_SIZE = 36
FONT_SIZE = 45
# 空格的大小，换行的时候也是要大小的不然，两行的距离太紧
FONT_BLANK_SIZE = 10
BEG_POINT = 5
BLACK_COLOR = 0 + 0 * 256 + 0 * 256 * 256

FONT_TYPE = 'C:\Windows\Fonts\FZSTK.TTF'  # 方正舒体
# FONT_TYPE = 'C:\Windows\Fonts\simsun.ttc'  # 常规简体   done 16 25
# FONT_TYPE = 'C:\Windows\Fonts\simhei.TTF'  # heiti     done 16 25
FONT_TYPE = 'C:\Windows\Fonts\STHUPO.TTF'  # 实心黑粗体  字体太大不好看有些字显示不正常
FONT_TYPE = 'C:\Windows\Fonts\simkai.TTF'  # 楷体 常规    done 16 25
FONT_TYPE = 'C:\Windows\Fonts\simfang.TTF'  # 仿宋 常规 常规  done 16 25
FONT_TYPE = 'C:\Windows\Fonts\FZYTK.TTF'    # 方正 姚体 常规
FONT_TYPE = 'C:\Windows\Fonts\STXINWEI.TTF'    # 华文新魏常规 done 25 36
FONT_TYPE = 'C:\Windows\Fonts\STCAIYUN.TTF'   #空心体
FONT_TYPE = 'F:\img_test\dahei\da_gei.TTF'    #自己下载的打黑 done  50
out_dir = 'F:/img_test/create_train_image/'
output_txt = 'output1.txt'
output_img = 'F:/img_test/create_train_image/output2.jpg'

# input_text = 'F:/img_test/create_train_image/common_word.txt'
input_text = 'F:/img_test/create_train_image/common_word2.txt'
input_text = 'F:/img_test/create_train_image/all_chinese.txt'
page_num = 47

def get_text_from_file():
with open(input_text, 'r') as file:
# decode('utf-8') 不行不知道为什么非要 'gbk'
# print file.read().decode('gbk')
return file.read().decode('gbk')

def get_page_num(text):
textLen = len(text)
rowLen = WIDTH / FONT_SIZE - 1
columnLen = HEIGHT / (FONT_SIZE + FONT_BLANK_SIZE)

if HEIGHT % (FONT_SIZE + FONT_BLANK_SIZE) > FONT_SIZE:
columnLen += 1

pageWordsNum = rowLen * columnLen
pageNum = float(textLen) / (pageWordsNum)

pageNum = int(round(pageNum + 0.499))

textLst = str_to_strlist_by_nth(text, pageWordsNum)

# for item in textLst:
#     print item

return pageNum, textLst

def do_All_Task():
text = get_text_from_file()
pageNum, textLst = get_page_num(text)

baseFileName = get_filename_without_extension(input_text)
baseFontName = get_filename_without_extension(FONT_TYPE)

for i in range(0, pageNum):
global output_txt, output_img, page_num
output_txt = '%s%d%s%s%spage%d%s%s%s%d.txt' % (out_dir, page_num, '_', baseFileName, '_', i, '_', baseFontName, '_', FONT_SIZE)
output_img = '%s%d%s%s%spage%d%s%s%s%d.jpg' % (out_dir, page_num, '_', baseFileName, '_', i, '_', baseFontName, '_', FONT_SIZE)
print (output_txt, output_img)
write_text2img(textLst[i])
page_num += 1

def get_filename_without_extension(fullPath):
base = os.path.basename(fullPath)
return os.path.splitext(base)[0]

def process_text(text):
# num of character in each line
row_size = WIDTH / FONT_SIZE - 1
# print row_len
LEN = len(text)
# print LEN
count = LEN / row_size

# for i in range(0, count):
#     text = insert_character(text, i * row_size + i)

textLst = str_to_strlist_by_nth(text, row_size)
# print text
return textLst

def write_text2img(text):
# img = np.zeros([WIDTH, HEIGHT, 3], dtype=np.uint8)
# img.fill(255) # or img[:] = 255
# cv2.imwrite('F:/img_test/create_train_image/1215.jpg', img)
# text = get_text_from_file()
text_lst = process_text(text)

blank = Image.new("RGB", [WIDTH, HEIGHT], "white")
# blank = Image.new("L", [WIDTH, HEIGHT], "white")
drawObject = ImageDraw.Draw(blank)
# font = ImageFont.truetype()
# Font1 = ImageFont.truetype("C:\Windows\Fonts\simsun.ttc", 36)
Font1 = ImageFont.truetype(FONT_TYPE, FONT_SIZE)
Font2 = ImageFont.truetype(FONT_TYPE, FONT_BLANK_SIZE)
# text = u"我草草草草\n\n草我草草草草"
# text = u"我草草草草草我草草草草"
drawObject.ink = BLACK_COLOR

lst_coord = []
dic_word2coord = {}
lst_word2coord = []

size = len(text_lst)
for i in range(0, size):
row_text = text_lst[i]
y = BEG_POINT + i * (FONT_SIZE + FONT_BLANK_SIZE)
drawObject.text([BEG_POINT, y], row_text, font=Font1)
drawObject.text([BEG_POINT, BEG_POINT + (i + 1) * FONT_SIZE + i * FONT_BLANK_SIZE], '\n', font=Font2)

for j in range(0, len(row_text)):
x = BEG_POINT + j * FONT_SIZE
# drawObject.rectangle((x, y, FONT_SIZE, FONT_SIZE + 2), outline="red")
lst_coord.append((x, y, x + FONT_SIZE, y + FONT_SIZE + 2))
dic_word2coord[row_text[j]] = (x, y, FONT_SIZE, FONT_SIZE + 2)
# lst_item = [row_text[j].encode('utf-8'), x, y, x + FONT_SIZE, y + FONT_SIZE + 2, 9]
lst_item = [row_text[j].encode('utf-8'), x, HEIGHT - y - FONT_SIZE - 2, x + FONT_SIZE, HEIGHT - y, page_num]
lst_word2coord.append(lst_item)

blank.save(output_img)

for j in range(0, len(lst_coord)):
drawObject.rectangle(lst_coord[j], outline="red")
# print dic_word2coord
blank.show()
# dict2txt(dic_word2coord)
lst2txt(lst_word2coord)

def lst2txt(lst):
filepath = output_txt
with open(filepath, "w") as f:
for item_lst in lst:
# f.write(item[0].encode('utf-8') + str(item[1]) + ' ' + str(item[2]) + ' ' + str(item[3]) + '\n')
# f.write(" ".join(item) + '\n')
f.write(" ".join(str(item) for item in item_lst) + '\n')

def dict2txt(d):
# txtfile = os.path.join(out_dir, output_txt)
filepath = 'F:/img_test/create_train_image/output1.txt'
with open(filepath, "w") as f:
for (k, v) in d.items():
k = k.encode('utf-8')
# f.write(k + v[0] + v[1] + v[2] + v[3])
f.write(k + ' ' + str(v[0]) + ' ' + str(v[1]) + ' ' + str(v[2]) + ' ' + str(v[3]) + '\n')
# for p in d.items():
#     f.write("%s:%s\n" % p)

def str_to_strlist_by_nth(str, nth):
return [str[i:i + nth] for i in range(0, len(str), nth)]

def insert_character(string, index):
return string[:index] + '\n' + string[index:]

def test():
text = get_text_from_file()
print process_text(text)

if __name__ == '__main__':
do_All_Task()

# print '%s%d%s%s.jpg' % (out_dir, 1, '_', 'abc')
# print(os.path.splitext(FONT_TYPE)[0])
# print(os.path.splitext(output_img)[0])
# base = os.path.basename(FONT_TYPE)
# print os.path.splitext(base)[0]
# write_text2img()
# print int(round(0.4))
# print int(round(0.1 + 0.49))
# test()
# d = {u'\u7981': (205, 425, 255, 477), u'\u5c04': (5, 365, 55, 417)}
# dict2txt(d)
# print len(get_text_from_file())
# str = u'1234我的5'
# print str_to_strlist_by_nth(str, 3)

代码在个人github