字体分类
本项目是PaddleOCR的其中一个环节。
目的是达到对PaddleOCR数据中图片的文字字体类别的识别。
0 汉字图片数据获取(可选)
需要增加字体类别时候再看此大步骤
本项目中以及放置了生成好的图片文件，不需要再执行此步骤,但下面的解压还是需要的，会用到里面的字体库
In [ ]
!unzip /home/aistudio/data/data122617/Createfont.zip -d work
记得自行修改两个文件中涉及到路径的内容，解压前的数据集的路径并不是正确的。
在work/Createfont/Word_Font_pre/下：
# -*- coding: utf-8 -*-
# 获取字体文件名，字体文件可以在百度下载或者电脑字体目录下寻找
import os
def Word_Font():
Word_Font_Path = 'work/Createfont/Word_Font_pre'
dirs = os.listdir(Word_Font_Path)
Word_Font_List = []
for dir in dirs:
Word_Font_List.append(dir)
return Word_Font_List在work/Createfont/下：
import re
f = open("work/Createfont/9004个常用汉字列表.txt", "r")
lines = f.readlines() # 读取全部内容 ，并以列表方式返回
Library = []
for line in lines:
line = line.split('\n')
line = re.sub(r'\n', "", line[0])
line = re.sub(':', "", line)
line = line.replace(" ", "")
line = line.lstrip('：')
line = line[:0] + line[13:]
for i in line:
Library.append(i)
size_L = len(Library)
# 主程序目录
from PIL import Image, ImageDraw, ImageFont, ImageOps
import os
import re
import Word_Font_pre.word_font as wf
# 读取字体文件
Word_Font_List = wf.Word_Font()
# 选择字体以及图片参数的初始值
class LetterImage():
def __init__(self, imgSize=(0, 0), imgMode='RGB', bg_color=(255, 255, 255), fg_color=(0, 0, 0),
fontsize=10, Word_Font=Word_Font_List[1]):
self.imgSize = imgSize
self.imgMode = imgMode
self.fontsize = fontsize
self.bg_color = bg_color
self.fg_color = fg_color
if(Word_Font.endswith("ttc")):
self.font = ImageFont.truetype(Word_Font, fontsize, index=1)
else:
self.font = ImageFont.truetype(Word_Font, fontsize)
# 设定生成图片大小
def GenLetterImage(self, letters):
self.letters = letters
(self.letterWidth, self.letterHeight) = self.font.getsize(letters)
if self.imgSize == (0, 0):
self.imgSize = (self.letterWidth - 0, self.letterHeight + 15) # 底边边距
self.imgWidth, self.imgHeight = self.imgSize
self.img = Image.new(self.imgMode, self.imgSize, self.bg_color)
self.drawBrush = ImageDraw.Draw(self.img)
textY0 = (self.imgHeight - self.letterHeight - 2)
textY0 = int(textY0)
textX0 = int((self.imgWidth - self.letterWidth - 2)) # 显示窗口坐标
self.drawBrush.text((textX0, textY0), self.letters, fill=self.fg_color, font=self.font)
if __name__ == '__main__':
f = open("work/Createfont/9004个常用汉字列表.txt", "r")
lines = f.readlines() # 读取全部内容 ，并以列表方式返回
Library = []
for line in lines:
line = line.split('\n')
line = re.sub(r'\n', "", line[0])
line = re.sub(':', "", line)
line = line.replace(" ", "")
line = line.lstrip('：')
# line = line[:0] + line[13:]
for i in line:
Library.append(i)
letterList = []
# ---------------将图片参数追加到列表以便后期调用--------------------
for j in range(0, len(Word_Font_List), 1):
try:
letterList.append(LetterImage(bg_color=(255, 255, 255), fontsize=100, Word_Font=Word_Font_List[j]))
finally:
print(Word_Font_List[j])
num_letter = len(Library) # 字体数量
# ---------------------------创建文件夹------------------------------
File_name = re.sub(r'\.', '_', Word_Font_List[j])
paths = os.getcwd()[:-4] + 'Images\\' + File_name # 获取此py文件路径，在此路径选创建文件夹
if not os.path.exists(paths):
os.makedirs(paths)
paths = paths + "\\"
# -----------------在某一种字体下，对字库遍历，生成相应字体图片------------------------
for i in range(num_letter - 1):
letterList[j].GenLetterImage(Library[i])
grayImg = ImageOps.grayscale(letterList[j].img)
grayImg.save(paths + str(i) + ".png")
In [ ]
# 若你想添加其他的类型字体
# 将字体的ttf或者ttc放入work/Createfont/Word_Font_pre下，直接执行下行代码即可
# !python work/Createfont/main.py
1 数据处理
- 制作标签文件label_list.txt
- 制作总的训练文件all_list.txt
- 打乱并划分为训练集合验证集
In [1]
!unzip -oq data/data122969/CreatenewImages.zip -d work
In [2]
# 当然不可避免的第一步，导入库
from sklearn.utils import shuffle
import os
from PIL import Image
import paddle
import random
In [3]
# -*- coding: utf-8 -*-
# 根据官方paddleclas的提示，我们需要把图像变为两个txt文件
# train_list.txt（训练集）
# val_list.txt（验证集）
# 根据左侧生成的文件夹名字来写根目录
# 训练数据只需要用到训练的，其他两个文件夹不需要读
dirpath = "work/CreatenewImages"
# 先得到总的txt后续再进行划分，因为要划分出验证集，所以要先打乱，因为原本是有序的
def get_all_txt():
all_list = []
label_list = []
i = 0 # 标记总文件数量
j = 0 # 标记文件类别
for root,dirs,files in os.walk(dirpath): # 分别代表根目录、文件夹、文件
# 遍历每个文件去生成all_list集合
for file in files:
# 文件中每行格式： 图像相对路径 图像的label_id（数字类别）（注意：中间有空格）。
imgpath = os.path.join(root,file)
# 有坏图，解决掉
try:
img = Image.open(imgpath)
except :
os.remove(imgpath)
else :
img = Image.open(imgpath)
img.convert('RGB')
i = i + 1
all_list.append(imgpath+" "+str(j-1)+"\n")
# 遍历每个dir去生成后续需要的label_list集合
m = 0
for d in dirs:
dpath = os.path.join(root,d)
d = dpath.split("/")[2]
label_list.append(d+"\n")
m = m+1
j = j + 1
# 生成all_list.txt
allstr = ''.join(all_list)
f = open('all_list.txt','w',encoding='utf-8')
f.write(allstr)
f.close
# 生成label_list.txt
labellist = ''.join(label_list)
f = open('label_list.txt','w',encoding='utf-8')
f.write(labellist)
f.close
return all_list , i
all_list,all_lenth = get_all_txt()
print(all_lenth)
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">216072
</span></span>
In [4]
# 把数据打乱
all_list = shuffle(all_list)
allstr = ''.join(all_list)
f = open('all_list.txt','w',encoding='utf-8')
f.write(allstr)
f.close()
print("打乱成功，并重新写入文本")
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">打乱成功，并重新写入文本
</span></span>
In [5]
# 按照比例划分数据集 数据有216072张图片，我这里采用9:1划分
train_size = int(all_lenth * 0.9)
train_list = all_list[:train_size]
val_list = all_list[train_size:]
print(len(train_list))
print(len(val_list))
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">194464
21608
</span></span>
In [6]
# 运行cell，生成训练集txt
train_txt = ''.join(train_list)
f_train = open('train_list.txt','w',encoding='utf-8')
f_train.write(train_txt)
f_train.close()
print("train_list.txt 生成成功！")
# 运行cell，生成验证集txt
val_txt = ''.join(val_list)
f_val = open('val_list.txt','w',encoding='utf-8')
f_val.write(val_txt)
f_val.close()
print("val_list.txt 生成成功！")
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">train_list.txt 生成成功！
val_list.txt 生成成功！
</span></span>
此时，如果正确，左侧有如图所示的文件
2 环境搭建
In [1]
! pip install paddlex
In [2]
import warnings
warnings.filterwarnings("ignore")
import paddlex as pdx
3 数据预处理
In [64]
from paddlex import transforms as T
train_transforms = T.Compose([
T.ResizeByShort(short_size=232, max_size=-1, interp='LINEAR'),
T.CenterCrop(crop_size=224),
T.RandomHorizontalFlip(),
T.RandomVerticalFlip(),
T.RandomBlur(prob=0.1),
T.Normalize()
])
eval_transforms = T.Compose([
T.ResizeByShort(),
T.CenterCrop(crop_size=224),
T.RandomHorizontalFlip(),
T.RandomVerticalFlip(),
T.RandomBlur(prob=0.1),
T.Normalize()
])
4 定义数据集dataset
In [10]
train_dataset = pdx.datasets.ImageNet(
data_dir='',
file_list='train_list.txt',
label_list='label_list.txt',
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.ImageNet(
data_dir='',
file_list='val_list.txt',
label_list='label_list.txt',
transforms=eval_transforms)
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">2021-12-23 17:46:38 [INFO] Starting to read file list from dataset...
2021-12-23 17:46:40 [INFO] 194464 samples in file train_list.txt
2021-12-23 17:46:40 [INFO] Starting to read file list from dataset...
2021-12-23 17:46:41 [INFO] 21608 samples in file val_list.txt
</span></span>
5 训练并验证
In [11]
num_classes = len(train_dataset.labels)
print(num_classes)
model = pdx.cls.ResNet50_vd_ssld(num_classes=num_classes)
model.train(num_epochs=5,
train_dataset=train_dataset,
train_batch_size=128,
eval_dataset=eval_dataset,
lr_decay_epochs=[6, 8],
save_interval_epochs=1,
learning_rate=0.00625,
save_dir='output/ResNet50_vd_ssld_bs128',
pretrain_weights=None,
use_vdl=True)
In [12]
model.evaluate(eval_dataset, batch_size=64, return_details=False)
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">2021-12-23 18:42:21 [INFO] Start to evaluate(total_samples=21608, total_steps=338)...
</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">OrderedDict([('acc1', 0.9433432), ('acc5', 0.9998613)])</span></span>
6 简单测试
In [3]
import paddlex as pdx
model = pdx.load_model('output/ResNet50_vd_ssld_bs128/best_model')
# 这张图片是华文琥珀的杨字
image_name = 'yang.png'
result = model.predict(image_name)
print("Predict Result:", result)
print(result[0]['category'])
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">W1230 19:41:37.423154 101 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W1230 19:41:37.428342 101 device_context.cc:465] device: 0, cuDNN Version: 7.6.
</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">2021-12-30 19:41:40 [INFO] Model[ResNet50_vd_ssld] loaded.
Predict Result: [{'category_id': 1, 'category': 'STHUPO_TTF', 'score': 0.9962198}]
STHUPO_TTF
</span></span>
7 PaddleOCR的应用
先了解一下基本的PaddleOCR是怎么用的
7.1 安装并测试demo
In [4]
!pip install "paddleocr>=2.0.1" --no-deps
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
Collecting paddleocr>=2.0.1
Downloading https://pypi.tuna.tsinghua.edu.cn/packages/e1/b6/5486e674ce096667dff247b58bf0fb789c2ce17a10e546c2686a2bb07aec/paddleocr-2.3.0.2-py3-none-any.whl (250kB)
|████████████████████████████████| 256kB 6.3MB/s eta 0:00:01
Installing collected packages: paddleocr
Successfully installed paddleocr-2.3.0.2
</span></span>
In [5]
!pip install shapely fasttext==0.9.1 imgaug==0.4.0 lmdb lxml premailer pyclipper python-Levenshtein scikit-image
In [6]
from paddleocr import PaddleOCR, draw_ocr
# Paddleocr目前支持中英文、英文、法语、德语、韩语、日语，可以通过修改lang参数进行切换
# 参数依次为`ch`, `en`, `french`, `german`, `korean`, `japan`。
ocr = PaddleOCR(use_angle_cls=True, lang="ch") # need to run only once to download and load model into memory
img_path = 'test.png'
result = ocr.ocr(img_path, cls=True)
for line in result:
print(line)
# 显示结果
from matplotlib import pyplot as plt
from PIL import Image
%matplotlib inline
image = Image.open(img_path).convert('RGB')
boxes = [line[0] for line in result]
txts = [line[1][0] for line in result]
scores = [line[1][1] for line in result]
im_show = draw_ocr(image, boxes, txts, scores, font_path='work/Createfont/Word_Font_pre/simfang.ttf')
im_show = Image.fromarray(im_show)
im_show.save('result.jpg')
plt.figure(figsize=(20,20))
plt.imshow(im_show)
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">[2021/12/30 19:43:00] root WARNING: version PP-OCRv2 not support cls models, auto switch to version PP-OCR
download https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar to /home/aistudio/.paddleocr/2.3.0.2/ocr/det/ch/ch_PP-OCRv2_det_infer/ch_PP-OCRv2_det_infer.tar
</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff"> 0%| | 0.00/3.19M [00:00<?, ?iB/s]</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">100%|██████████| 3.19M/3.19M [00:00<00:00, 8.46MiB/s]
</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">download https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer.tar to /home/aistudio/.paddleocr/2.3.0.2/ocr/rec/ch/ch_PP-OCRv2_rec_infer/ch_PP-OCRv2_rec_infer.tar
</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff"> 0%| | 0.00/8.88M [00:00<?, ?iB/s]</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">100%|██████████| 8.88M/8.88M [00:01<00:00, 7.90MiB/s]
</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">download https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar to /home/aistudio/.paddleocr/2.3.0.2/ocr/cls/ch_ppocr_mobile_v2.0_cls_infer/ch_ppocr_mobile_v2.0_cls_infer.tar
</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff"> 0%| | 0.00/1.45M [00:00<?, ?iB/s]</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">100%|██████████| 1.45M/1.45M [00:00<00:00, 4.33MiB/s]
</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">Namespace(benchmark=False, cls_batch_num=6, cls_image_shape='3, 48, 192', cls_model_dir='/home/aistudio/.paddleocr/2.3.0.2/ocr/cls/ch_ppocr_mobile_v2.0_cls_infer', cls_thresh=0.9, cpu_threads=10, det=True, det_algorithm='DB', det_db_box_thresh=0.6, det_db_score_mode='fast', det_db_thresh=0.3, det_db_unclip_ratio=1.5, det_east_cover_thresh=0.1, det_east_nms_thresh=0.2, det_east_score_thresh=0.8, det_limit_side_len=960, det_limit_type='max', det_model_dir='/home/aistudio/.paddleocr/2.3.0.2/ocr/det/ch/ch_PP-OCRv2_det_infer', det_pse_box_thresh=0.85, det_pse_box_type='box', det_pse_min_area=16, det_pse_scale=1, det_pse_thresh=0, det_sast_nms_thresh=0.2, det_sast_polygon=False, det_sast_score_thresh=0.5, drop_score=0.5, e2e_algorithm='PGNet', e2e_char_dict_path='./ppocr/utils/ic15_dict.txt', e2e_limit_side_len=768, e2e_limit_type='max', e2e_model_dir=None, e2e_pgnet_mode='fast', e2e_pgnet_polygon=True, e2e_pgnet_score_thresh=0.5, e2e_pgnet_valid_set='totaltext', enable_mkldnn=False, gpu_mem=500, help='==SUPPRESS==', image_dir=None, ir_optim=True, label_list=['0', '180'], lang='ch', layout_path_model='lp://PubLayNet/ppyolov2_r50vd_dcn_365e_publaynet/config', max_batch_size=10, max_text_length=25, min_subgraph_size=15, ocr_version='PP-OCRv2', output='./output/table', precision='fp32', process_id=0, rec=True, rec_algorithm='CRNN', rec_batch_num=6, rec_char_dict_path='/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddleocr/ppocr/utils/ppocr_keys_v1.txt', rec_image_shape='3, 32, 320', rec_model_dir='/home/aistudio/.paddleocr/2.3.0.2/ocr/rec/ch/ch_PP-OCRv2_rec_infer', save_log_path='./log_output/', show_log=True, structure_version='STRUCTURE', table_char_dict_path=None, table_char_type='en', table_max_len=488, table_model_dir=None, total_process_num=1, type='ocr', use_angle_cls=True, use_dilation=False, use_gpu=True, use_mp=False, use_onnx=False, use_pdserving=False, use_space_char=True, use_tensorrt=False, vis_font_path='./doc/fonts/simfang.ttf', warmup=True)
[2021/12/30 19:43:04] root DEBUG: dt_boxes num : 2, elapse : 0.023739099502563477
[2021/12/30 19:43:04] root DEBUG: cls num : 2, elapse : 0.009242057800292969
[2021/12/30 19:43:04] root DEBUG: rec_res num : 2, elapse : 0.008728265762329102
[[[84.0, 66.0], [463.0, 54.0], [465.0, 125.0], [87.0, 137.0]], ('幸福通道', 0.9984914)]
[[[247.0, 491.0], [395.0, 493.0], [394.0, 530.0], [246.0, 527.0]], ('禁止入内', 0.9945936)]
</span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff"><matplotlib.image.AxesImage at 0x7fbc78407c50></span></span>
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff"><Figure size 1440x1440 with 1 Axes></span></span>
7.2 处理字体
思路：
- 把框起来的区域截取作为一张图片
- 截取的这张图片进入上面的分类模型进行预测，得到字体的分类结果
- 按照PaddleOCR的规则进行输出展示
7.3 导入训练好的模型
In [7]
import paddlex as pdx
model = pdx.load_model('output/ResNet50_vd_ssld_bs128/best_model')
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">2021-12-30 19:43:15 [INFO] Model[ResNet50_vd_ssld] loaded.
</span></span>
7.4 ☆ 处理输出
截取片段，然后识别
In [17]
import os
# 显示结果
from matplotlib import pyplot as plt
from PIL import Image
%matplotlib inline
def jietu_smallpart(imgSrc,result):
IMG = imgSrc # 图片地址
im = Image.open(IMG) # 用PIL打开一个图片
im = im.convert("RGB")
ziti_list=[]
zitip_list=[]
i = 0
for line in result:
# print(line[0][0][0], line[0][0][1], line[0][2][0], line[0][2][1])
# box = (line[0][0][0]-2, line[0][0][1]-2, line[0][2][0]+2, line[0][2][1]+2)
# box = (line[0][0][0]-2, line[0][0][1]-2, line[0][0][0] + line[0][2][1] - line[0][0][1], line[0][2][1])
box = (line[0][0][0]-10, line[0][0][1]-10, line[0][2][0]+10, line[0][2][1]+10)
# box代表需要剪切图片的位置格式为:xmin ymin xmax ymax
ng = im.crop(box) # 对im进行裁剪 保存为ng(这里im保持不变)
# ng = ng.resize((58,58))
image_name = str(i) + '_copy.jpg'
i = i+1
print(image_name)
ng.save(image_name)
# 拿到这张图
result = model.predict(image_name)
ziti_list.append(result[0]['category'])
zitip_list.append(result[0]['score'])
return ziti_list,zitip_list
test,testp = jietu_smallpart("test.png",result)
print(test)
print(testp)
<span style="color:rgba(0, 0, 0, 0.85)"><span style="background-color:#ffffff">0_copy.jpg
1_copy.jpg
['simsun_ttc', 'simsun_ttc']
[0.999925, 0.8224375]
</span></span>
总结
- 训练字体的分类，但是目前来看虽然验证集精度很高，在实际的场景中运用，问题还是存在的。
- 还没有进行模块化函数式的编程
- 后续需要再想办法提升模型的鲁棒性
本文章为转载内容，我们尊重原作者对文章享有的著作权。如有内容错误或侵权问题，欢迎原作者联系我们进行内容更正或删除文章。
