EasyOCR 识别模型训练

2.1、项目安装

$ git clone https://github.com/DaveLogs/TRDG2DTRB.git

2.2、数据转换

python3 convert.py  --input_path /home/ocr/  --output_path ./output

原始图片的命名是有要求的：图片内容_index编号.后缀

像4051.jpg这种格式的经过转换后得到的gt.txt如下，不是我们想要的

相关代码逻辑如下：

3.1、项目安装

# 下载源代码
$ git clone https://github.com/clovaai/deep-text-recognition-benchmark.git
 
# 搭建开发环境
$ pip3 install torch torchvision
$ pip3 install lmdb pillow nltk natsort
$ pip3 install fire

3.2.1、训练数据

3.2.2、将训练数据转换为lmdb格式

# deep-text-recognition-benchmark 从项目根运行
(venv) $ python3 create_lmdb_dataset.py \
        --inputPath /home/TRDG2DTRB/output/ \
        --gtFile /home/TRDG2DTRB/output/gt.txt \
        --outputPath result/

3.2.3、准备预训练模型

下载学习模型 跳转中...https://link.zhihu.com/?target=https%3A//github.com/clovaai/deep-text-recognition-benchmark%23run-demo-with-pretrained-model 下载与实际 EasyOCR 中使用的基本模型具有相同网络结构（' None-VGG-BiLSTM-CTC '）的预训练模型。

3.2.4、项目和预模型正常运行的确认

# demo.py中可查看参数及其定义

 

python3 demo.py \
--Transformation None \
--FeatureExtraction VGG \
--SequenceModeling BiLSTM \
--Prediction CTC \
--image_folder demo_image/ \
--saved_model None-VGG-BiLSTM-CTC.pth

3.3、训练模型

# train.py中查看参数及其定义

 

python3 train.py --train_data lmdb/training \
--valid_data lmdb/validation \
--select_data MJ-ST \
--batch_ratio 0.5-0.5 \
--Transformation None \
--FeatureExtraction VGG \
--SequenceModeling BiLSTM \
--Prediction CTC \
--saved_model None-VGG-BiLSTM-CTC.pth \
--num_iter 2000 \
--valInterval 20 \
--FT

上述命令语法的简要说明如下。

• --train_data : 训练数据中训练的数据路径
• --valid_data : 训练数据之间验证的数据路径
• --select_data : 选择训练数据（默认为MJ-ST，即MJ和ST作为训练数据）
• --batch_ratio：为批次中的每个选定数据分配比率
• --Transformation：选择要使用的转换模块。['无'，'TPS']
• --FeatureExtraction : 选择要使用的 FeatureExtraction 模块,['RCNN'、'ResNet'、'VGG']
• --SequenceModeling：选择要使用的 SequenceModeling 模块。['无'，'BiLSTM']
• --Prediction：选择要使用的预测模块。['Attn', 'CTC']
• --saved_model : 用于微调学习的预训练模型的存储位置
• --num_iter: 训练迭代次数，默认300000
• --valInterval:每次检验之间的时间间隔，默认2000
• --FT : 是否学习微调
• --lr：学习率，对于 Adadelta，默认 = 1.0
• --batch_max_length：最大标签长度，默认值25
• --imgH：输入图像的高度，默认32      # 后面的识别配置模块nvbc.yaml文件会用到
• --input_channel：特征提取器的输入通道数，默认1
• --output_channel：特征提取器的输出通道数，默认512
• --hidden_size：LSTM 隐藏状态的大小，默认256

• best_accuracy.pth / best_norm_ED.pth：在经过训练的模型文件中具有特定性能指数的选定模型；
• log_dataset.txt：用于训练的数据集信息；
• log_train.txt：训练正在进行时的日志（与上面终端中显示的相同）
• opt.txt：执行学习命令语法时设置的学习选项信息

3.4、测试模型

# 测试项目中包含的演示图像
python3 demo.py \
--Transformation None \
--FeatureExtraction VGG \
--SequenceModeling BiLSTM \
--Prediction CTC \
--image_folder demo_image/ \
--saved_model ./saved_models/None-VGG-BiLSTM-CTC-Seed1111/best_accuracy.pth

network_params:
  input_channel: 1
  output_channel: 512
  hidden_size: 256
imgH: 32
lang_list:
         - 'nvbc'   # 语言代码   对应与/usr/local/lib/python3.6/dist-packages/easyocr/character/nvbc_char.txt，没有则创建
character_list: 0123456789abcdefghijklmnopqrstuvwxyz   # 学习数据类

import torch.nn as nn
 
class Model(nn.Module):
    def __init__(self, input_channel, output_channel, hidden_size, num_class):
        super(Model, self).__init__()
        """ FeatureExtraction """
        self.FeatureExtraction = VGG_FeatureExtractor(input_channel, output_channel)
        self.FeatureExtraction_output = output_channel
        self.AdaptiveAvgPool = nn.AdaptiveAvgPool2d((None, 1))
 
        """ Sequence modeling"""
        self.SequenceModeling = nn.Sequential(
            BidirectionalLSTM(self.FeatureExtraction_output, hidden_size, hidden_size),
            BidirectionalLSTM(hidden_size, hidden_size, hidden_size))
        self.SequenceModeling_output = hidden_size
 
        """ Prediction """
        self.Prediction = nn.Linear(self.SequenceModeling_output, num_class)
 
    def forward(self, input, text):
        """ Feature extraction stage """
        visual_feature = self.FeatureExtraction(input)
        visual_feature = self.AdaptiveAvgPool(visual_feature.permute(0, 3, 1, 2))
        visual_feature = visual_feature.squeeze(3)
 
        """ Sequence modeling stage """
        contextual_feature = self.SequenceModeling(visual_feature)
 
        """ Prediction stage """
        prediction = self.Prediction(contextual_feature.contiguous())
 
        return prediction
 
class BidirectionalLSTM(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super(BidirectionalLSTM, self).__init__()
        self.rnn = nn.LSTM(input_size, hidden_size, bidirectional=True, batch_first=True)
        self.linear = nn.Linear(hidden_size * 2, output_size)
 
    def forward(self, input):
        """
        input : visual feature [batch_size x T x input_size]
        output : contextual feature [batch_size x T x output_size]
        """
        try: # multi gpu needs this
            self.rnn.flatten_parameters()
        except: # quantization doesn't work with this
            pass
        recurrent, _ = self.rnn(input)  # batch_size x T x input_size -> batch_size x T x (2*hidden_size)
        output = self.linear(recurrent)  # batch_size x T x output_size
        return output
 
 
class VGG_FeatureExtractor(nn.Module):
    def __init__(self, input_channel, output_channel=256):
        super(VGG_FeatureExtractor, self).__init__()
        self.output_channel = [int(output_channel / 8), int(output_channel / 4),
                               int(output_channel / 2), output_channel]
        self.ConvNet = nn.Sequential(
            nn.Conv2d(input_channel, self.output_channel[0], 3, 1, 1), nn.ReLU(True),
            nn.MaxPool2d(2, 2),
            nn.Conv2d(self.output_channel[0], self.output_channel[1], 3, 1, 1), nn.ReLU(True),
            nn.MaxPool2d(2, 2),
            nn.Conv2d(self.output_channel[1], self.output_channel[2], 3, 1, 1), nn.ReLU(True),
            nn.Conv2d(self.output_channel[2], self.output_channel[2], 3, 1, 1), nn.ReLU(True),
            nn.MaxPool2d((2, 1), (2, 1)),
            nn.Conv2d(self.output_channel[2], self.output_channel[3], 3, 1, 1, bias=False),
            nn.BatchNorm2d(self.output_channel[3]), nn.ReLU(True),
            nn.Conv2d(self.output_channel[3], self.output_channel[3], 3, 1, 1, bias=False),
            nn.BatchNorm2d(self.output_channel[3]), nn.ReLU(True),
            nn.MaxPool2d((2, 1), (2, 1)),
            nn.Conv2d(self.output_channel[3], self.output_channel[3], 2, 1, 0), nn.ReLU(True))
 
    def forward(self, input):
        return self.ConvNet(input)

from easyocr.easyocr import *
 
# # GPU 环境
# os.environ['CUDA_VISIBLE_DEVICES'] = '0,1'
 
def get_files(path):
    files = [f for f in os.listdir(path) if not f.startswith('.')]  # skip hidden file
    files.sort()
    abspath = os.path.abspath(path)
 
    file_list = []
    for file in files:
        file_path = os.path.join(abspath, file)
        file_list.append(file_path)
 
    return file_list, len(file_list)
 
if __name__ == '__main__':
    # Using custom model
    reader = Reader(['nvbc'], gpu=False,   # 语言存储在/usr/local/lib/python3.6/dist-packages/easyocr/character/nvbc_char.txt
                    model_storage_directory='/root/.EasyOCR/model',  
                    user_network_directory='/root/.EasyOCR/user_network',
                    recog_network='nvbc')
 
    files, count = get_files(path='/home/deep-text-recognition-benchmark/demo_image/')
    for idx, file in enumerate(files):
        filename = os.path.basename(file)
        result = reader.readtext(file)
        # ./easyocr/utils.py 733 lines
        # result[0]: bbox
        # result[1]: string
        # result[2]: confidence
        for (bbox, string, confidence) in result:
            print("filename: '%s', confidence: %.4f, string: '%s'" % (filename, confidence, string))

使用用户模型运行： python3 testzq.py，结果如下：

​