• 笔记记录--基于ccpd数据集利用Paddle OCR训练车牌检测模型

    1-- 环境搭建

    安装Paddle OCR参考

    ① 创建环境

    1. conda create -n paddle_env python=3.8
    2. conda activate paddle_env

    ② 安装paddlepaddle
    # 切换cuda版本为11.1（根据个人实际修改）

    1. sudo gedit ~/.bashrc
    2. source ~/.bashrc

    # 安装paddlepaddle

    python -m pip install paddlepaddle-gpu==2.3.0.post111 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html 

    ③ 安装依赖
    # 安装PaddleOCR whl

    pip install "paddleocr>=2.0.1"

    # 版面分析

    pip3 install -U https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl

    2 -- 数据集划分

    # 采用ccpd数据集的challenge系列
    # 45003用于训练集，5000用于验证集
    数据集文件转移代码

    1. import shutil
    2. import os
    3.  
    4. def remove_file(old_path, new_path):
    5.     filelist = os.listdir(old_path)  # 列出该目录下的所有文件,listdir返回的文件列表是不包含路径的。
    6.     i = 0
    7.     for file in filelist:
    8.         src = os.path.join(old_path, file)
    9.         dst = os.path.join(new_path, file)
    10.         
    11.         if i < 5000:
    12.             shutil.move(src, dst)
    13.         else:
    14.             break
    15.         i = i + 1
    16.         
    17. if __name__ == '__main__':
    18.     remove_file(r"/civi/Chinese_license_plate_Note/detection/dataset/ccpd_challenge", r"/civi/Chinese_license_plate_Note/detection/dataset/test_dataset")

    # 创建label标注文件

    格式参考

     代码样例

    1. import os
    2.  
    3. words_list = [
    4.     "A", "B", "C", "D", "E",
    5.     "F", "G", "H", "J", "K",
    6.     "L", "M", "N", "P", "Q",
    7.     "R", "S", "T", "U", "V",
    8.     "W", "X", "Y", "Z", "0",
    9.     "1", "2", "3", "4", "5",
    10.     "6", "7", "8", "9"
    11. ]
    12.  
    13. con_list = [
    14.     "皖", "沪", "津", "渝", "冀",
    15.     "晋", "蒙", "辽", "吉", "黑",
    16.     "苏", "浙", "京", "闽", "赣",
    17.     "鲁", "豫", "鄂", "湘", "粤",
    18.     "桂", "琼", "川", "贵", "云",
    19.     "西", "陕", "甘", "青", "宁",
    20.     "新"
    21. ]
    22.  
    23. if __name__ == "__main__":
    24.     points = []
    25.     label = []
    26.     for item in os.listdir(os.path.join('/civi/Chinese_license_plate_Note/detection/dataset/test_dataset/')): # 遍历图片
    27.  
    28.         _, _, bbox, points, label, _, _ = item.split('-') # 分割文件名
    29.  
    30.         points = points.split('_') # 分割四个坐标点
    31.         tmp = points
    32.         points = []
    33.         for _ in tmp:
    34.             points.append([int(_.split('&')[0]), int(_.split('&')[1])])
    35.         # print(points)
    36.  
    37.         label = label.split('_')
    38.         con = con_list[int(label[0])]
    39.         words = [words_list[int(_)] for _ in label[1:]]
    40.         label = con + ''.join(words)
    41.         label = '"' + label + '"'
    42.         file_name = item
    43.         List = '[{"transcription": ' + label + ', "points": ' + str(points) + '}]'
    44.         line = file_name + '\t' + List + '\n'
    45.  
    46.         with open('/civi/Chinese_license_plate_Note/detection/dataset/' + 'test_label.txt', 'a', encoding='UTF-8') as f:
    47.             f.write(line)

    上述代码博主犯了一个错误，就是CCPD数据集的四个坐标是从右下坐标顺时针开始的，而OCR检测的标注文件，其坐标要求从左上顺时针开始，所以上述代码修改为：

    1. import os
    2.  
    3. words_list = [
    4.     "A", "B", "C", "D", "E",
    5.     "F", "G", "H", "J", "K",
    6.     "L", "M", "N", "P", "Q",
    7.     "R", "S", "T", "U", "V",
    8.     "W", "X", "Y", "Z", "0",
    9.     "1", "2", "3", "4", "5",
    10.     "6", "7", "8", "9"
    11. ]
    12.  
    13. con_list = [
    14.     "皖", "沪", "津", "渝", "冀",
    15.     "晋", "蒙", "辽", "吉", "黑",
    16.     "苏", "浙", "京", "闽", "赣",
    17.     "鲁", "豫", "鄂", "湘", "粤",
    18.     "桂", "琼", "川", "贵", "云",
    19.     "西", "陕", "甘", "青", "宁",
    20.     "新"
    21. ]
    22.  
    23. if __name__ == "__main__":
    24.     points = []
    25.     label = []
    26.     for item in os.listdir(os.path.join('/civi/Chinese_license_plate_Note/detection/dataset/test_dataset/')): # 遍历图片
    27.  
    28.         _, _, bbox, points, label, _, _ = item.split('-') # 分割文件名
    29.  
    30.         points = points.split('_') # 分割四个坐标点
    31.         tmp = points
    32.         points = []
    33.         points.append([int(tmp[2].split('&')[0]), int(tmp[2].split('&')[1])])
    34.         points.append([int(tmp[3].split('&')[0]), int(tmp[3].split('&')[1])])
    35.         points.append([int(tmp[0].split('&')[0]), int(tmp[0].split('&')[1])])
    36.         points.append([int(tmp[1].split('&')[0]), int(tmp[1].split('&')[1])])
    37.  
    38.         label = label.split('_')
    39.         con = con_list[int(label[0])]
    40.         words = [words_list[int(_)] for _ in label[1:]]
    41.         label = con + ''.join(words)
    42.         label = '"' + label + '"'
    43.         file_name = item
    44.         List = '[{"transcription": ' + label + ', "points": ' + str(points) + '}]'
    45.         line = file_name + '\t' + List + '\n'
    46.  
    47.         with open('/civi/Chinese_license_plate_Note/detection/dataset/' + 'test_label.txt', 'a', encoding='UTF-8') as f:
    48.             f.write(line)

    3-- 训练模型

    训练参考

    ① 下载预训练模型（DB-Net）

     下载地址

    ② 配置config文件

    示例：（注释部分需要留意并修改）

    1. Global:
    2.   use_gpu: true # 是否使用gpu
    3.   epoch_num: 200  # epoch数目
    4.   log_smooth_window: 20
    5.   print_batch_step: 2
    6.   save_model_dir: /civi/Chinese_license_plate_Note/detection/Models_Well_trained/200epochs # 保存模型的地址
    7.   save_epoch_step: 10 # 保存模型的间隔
    8.   # evaluation is run every 5000 iterations after the 4000th iteration
    9.   eval_batch_step: [3000, 2000]
    10.   cal_metric_during_train: False
    11.   pretrained_model: /civi/Chinese_license_plate_Note/detection/pretrain/ch_ppocr_server_v2.0_det_train/best_accuracy  # 预训练模型的地址
    12.   checkpoints:
    13.   save_inference_dir:
    14.   use_visualdl: False
    15.   infer_img: /civi/Chinese_license_plate_Note/detection/test_img/test6.22.png # 测试图片
    16.   save_res_path: /civi/Chinese_license_plate_Note/detection/Models_Well_trained/100epochs/det_db/predicts_db.txt
    17.  
    18. Architecture:
    19.   model_type: det
    20.   algorithm: DB
    21.   Transform:
    22.   Backbone:
    23.     name: ResNet
    24.     layers: 18
    25.     disable_se: True
    26.   Neck:
    27.     name: DBFPN
    28.     out_channels: 256
    29.   Head:
    30.     name: DBHead
    31.     k: 50
    32.  
    33. Loss:
    34.   name: DBLoss
    35.   balance_loss: true
    36.   main_loss_type: DiceLoss
    37.   alpha: 5
    38.   beta: 10
    39.   ohem_ratio: 3
    40.  
    41. Optimizer:
    42.   name: Adam
    43.   beta1: 0.9
    44.   beta2: 0.999
    45.   lr:
    46.     name: Cosine
    47.     learning_rate: 0.001
    48.     warmup_epoch: 2
    49.   regularizer:
    50.     name: 'L2'
    51.     factor: 0
    52.  
    53. PostProcess:
    54.   name: DBPostProcess
    55.   thresh: 0.3
    56.   box_thresh: 0.6
    57.   max_candidates: 1000
    58.   unclip_ratio: 1.5
    59.  
    60. Metric:
    61.   name: DetMetric
    62.   main_indicator: hmean
    63.  
    64. Train:
    65.   dataset:
    66.     name: SimpleDataSet
    67.     data_dir: /civi/Chinese_license_plate_Note/detection/dataset/train_dataset/ # 训练集图片
    68.     label_file_list:
    69.       - /civi/Chinese_license_plate_Note/detection/dataset/train_label.txt # 训练集标签
    70.     ratio_list: [1.0]
    71.     transforms:
    72.       - DecodeImage: # load image
    73.           img_mode: BGR
    74.           channel_first: False
    75.       - DetLabelEncode: # Class handling label
    76.       - IaaAugment:
    77.           augmenter_args:
    78.             - { 'type': Fliplr, 'args': { 'p': 0.5 } }
    79.             - { 'type': Affine, 'args': { 'rotate': [-10, 10] } }
    80.             - { 'type': Resize, 'args': { 'size': [0.5, 3] } }
    81.       - EastRandomCropData:
    82.           size: [960, 960]
    83.           max_tries: 50
    84.           keep_ratio: true
    85.       - MakeBorderMap:
    86.           shrink_ratio: 0.4
    87.           thresh_min: 0.3
    88.           thresh_max: 0.7
    89.       - MakeShrinkMap:
    90.           shrink_ratio: 0.4
    91.           min_text_size: 8
    92.       - NormalizeImage:
    93.           scale: 1./255.
    94.           mean: [0.485, 0.456, 0.406]
    95.           std: [0.229, 0.224, 0.225]
    96.           order: 'hwc'
    97.       - ToCHWImage:
    98.       - KeepKeys:
    99.           keep_keys: ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask'] # the order of the dataloader list
    100.   loader:
    101.     shuffle: True
    102.     drop_last: False
    103.     batch_size_per_card: 8 # batchsize
    104.     num_workers: 4
    105.  
    106. Eval:
    107.   dataset:
    108.     name: SimpleDataSet
    109.     data_dir: /civi/Chinese_license_plate_Note/detection/dataset/test_dataset/ # 验证集图片
    110.     label_file_list:
    111.       - /civi/Chinese_license_plate_Note/detection/dataset/test_label.txt # 验证集标签
    112.     transforms:
    113.       - DecodeImage: # load image
    114.           img_mode: BGR
    115.           channel_first: False
    116.       - DetLabelEncode: # Class handling label
    117.       - DetResizeForTest:
    118. #           image_shape: [736, 1280]
    119.       - NormalizeImage:
    120.           scale: 1./255.
    121.           mean: [0.485, 0.456, 0.406]
    122.           std: [0.229, 0.224, 0.225]
    123.           order: 'hwc'
    124.       - ToCHWImage:
    125.       - KeepKeys:
    126.           keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
    127.   loader:
    128.     shuffle: False
    129.     drop_last: False
    130.     batch_size_per_card: 1 # must be 1
    131.     num_workers: 2

    ③ 开始训练

    # 这里博主使用多卡训练，gpu编号为1和2

    1. python3 -m paddle.distributed.launch --gpus '1,2' tools/train.py \
    2. -c /civi/Chinese_license_plate_Note/detection/test.yml

    ## 未完待续

  • 相关阅读:
    Mybatis Plus配置多个数据源
    JAVA基础（十四）
    2022杭电多校赛第八场
    关于GIS空间分析的几点思路
    jsp获取数据 jsp直接获取后端数据 获取input选中的值 单选 没 checked属性
    【python】爬虫系列Day03--url传参
    Bootstrap的CSS类积累学习
    vue3-基础知识（4）- 组件
    [计算机入门] Windows功能的安装与卸载
    SVN自动更新
  • 原文地址：https://blog.csdn.net/weixin_43863869/article/details/125414476