OCR 检测数据集详解

前言
所有数据集目录组织形式如下所示(以ICDAR2015为例)

cd icdar2015

目录组织形式

|-- results
|   |-- result_0.jpg
|-- train
|   |-- gt
|   |   |-- CIMG0005_convert.txt
|   `-- img
|       |-- CIMG0005_convert.jpg
`-- train.txt

• ICDAR2015
  • Step1 数据集下载
    https://rrc.cvc.uab.es/?ch=4&com=downloads

    wget https://rrc.cvc.uab.es/downloads/ch4_training_images.zip
    wget https://rrc.cvc.uab.es/downloads/ch4_training_localization_transcription_gt.zip  
    wget https://rrc.cvc.uab.es/downloads/ch4_test_images.zip  
    wget https://rrc.cvc.uab.es/downloads/Challenge4_Test_Task1_GT.zip 
    

  • Step2 数据集规整
    按前言的目录组织，将下载的文件拷贝到对应文件目录
  • Step3 产生数据索引文件train.txt、test.txt

    from __future__ import absolute_import
    from __future__ import division
    from __future__ import print_function
    from __future__ import unicode_literals
    
    import os, sys
    import numpy as np
    
    icdar2015_root_dir = os.getcwd()
    icdar2015_train_img_dir = os.path.join(icdar2015_root_dir, 'train', 'img')
    icdar2015_train_gt_dir = os.path.join(icdar2015_root_dir, 'train', 'gt')
    
    icdar2015_test_img_dir = os.path.join(icdar2015_root_dir, 'test', 'img')
    icdar2015_test_gt_dir = os.path.join(icdar2015_root_dir, 'test', 'gt')
    
    print(f'icdar2015_root_dir:{icdar2015_root_dir}')
    print(f'icdar2015_train_img_dir:{icdar2015_train_img_dir}')
    print(f'icdar2015_train_gt_dir:{icdar2015_train_gt_dir}')
    
    print(f'icdar2015_test_img_dir:{icdar2015_test_img_dir}')
    print(f'icdar2015_test_gt_dir:{icdar2015_test_gt_dir}')
    print('*'*80)
    
    with open('train.txt', 'w') as f:
        imgs = os.listdir(icdar2015_train_img_dir)
        print(f'write train images:{len(imgs)}')
        for img in imgs:
            img_path = os.path.join(icdar2015_train_img_dir, img)
            gt_name = 'gt_' + img.replace('jpg', 'txt')
            gt_path = os.path.join(icdar2015_train_gt_dir, gt_name)
            f.write(img_path + '\t' + gt_path + '\n')
    
    with open('test.txt', 'w') as f:
        imgs = os.listdir(icdar2015_test_img_dir)
        print(f'write test images:{len(imgs)}')
        for img in imgs:
            img_path = os.path.join(icdar2015_test_img_dir, img)
            gt_name = 'gt_' + img.replace('jpg', 'txt')
            gt_path = os.path.join(icdar2015_test_gt_dir, gt_name)
            f.write(img_path + '\t' + gt_path + '\n')        
            
    print(f'****************** generate label list done ******************')  
    

  • Step4 测试

    #!/usr/bin/env python
    # coding=utf-8
    """
    @Autor: xinyi61
    @Date: 2020-02-23 14:32:59
    @LastEditors: xinyi61
    @LastEditTime: 2020-02-24 16:22:44
    @Email: [email protected]
    @Version: 1.0
    @Description: 
    """
    from __future__ import absolute_import
    from __future__ import division
    from __future__ import print_function
    from __future__ import unicode_literals
    
    import os, sys
    import random
    import numpy as np
    import matplotlib.pyplot as plt
    import cv2
    
    def get_annotation(label_path: str) -> tuple:
        boxes = []
        text_tags = []
        with open(label_path, encoding="utf-8", mode="r") as f:
            for line in f.readlines():
                params = line.strip().strip("\ufeff").strip("\xef\xbb\xbf").split(",")
                try:
                    label = params[8]
                    if label == "*" or label == "###":
                        text_tags.append(False)
                    else:
                        text_tags.append(True)
    
                    x1, y1, x2, y2, x3, y3, x4, y4 = list(map(float, params[:8]))
                    boxes.append([[x1, y1], [x2, y2], [x3, y3], [x4, y4]])
                except:
                    print("load label failed on {}".format(label_path))
    
        return np.array(boxes, dtype=np.float32), np.array(text_tags, dtype=np.bool)
    
    
    if __name__ == "__main__":
        color = [255, 0, 0]
        thickness = 1
    
        result = []
        with open("train.txt", "r") as f:
            lines = f.readlines()
            random.shuffle(lines)
            for line in lines[0:10]:
                line_info = line.rstrip("\n").split("\t")
                image_path = line_info[0]
                label_path = line_info[1]
                points, text_tags = get_annotation(label_path)
                img = cv2.imread(image_path)
                for i in range(len(points)):
                    if text_tags[i] == False:
                        continue
    
                    point = points[i]
                    point = point.astype(int)
                    cv2.line(img, tuple(point[0]), tuple(point[1]), color, thickness)
                    cv2.line(img, tuple(point[1]), tuple(point[2]), color, thickness)
                    cv2.line(img, tuple(point[2]), tuple(point[3]), color, thickness)
                    cv2.line(img, tuple(point[3]), tuple(point[0]), color, thickness)
    
                result.append(img[:, :, ::-1])
    
        i = 1
        for k in range(len(result)):
            try:
                plt.ion()
                plt.figure(i)
                cv2.imwrite("./results/result_{}.jpg".format(k), result[k])
                plt.imshow(result[k][:, :, ::-1])
                plt.pause(1)
            except:
                pass
            finally:
                i += 1
    
        print(f"************* done ***************")
    

• ICDAR2017
  • Step1 数据集下载

    wget http://datasets.cvc.uab.es/rrc/ch8_training_images_1.zip  
    wget http://datasets.cvc.uab.es/rrc/ch8_training_images_2.zip   
    wget http://datasets.cvc.uab.es/rrc/ch8_training_images_3.zip  
    wget http://datasets.cvc.uab.es/rrc/ch8_training_images_4.zip   
    wget http://datasets.cvc.uab.es/rrc/ch8_training_images_5.zip  
    wget http://datasets.cvc.uab.es/rrc/ch8_training_images_6.zip  
    wget http://datasets.cvc.uab.es/rrc/ch8_training_images_7.zip  
    wget http://datasets.cvc.uab.es/rrc/ch8_training_images_8.zip  
    
    wget http://datasets.cvc.uab.es/rrc/ch8_training_localization_transcription_gt_v2.zip   
    wget https://rrc.cvc.uab.es/downloads/ch8_validation_images.zip
    wget http://datasets.cvc.uab.es/rrc/ch8_validation_localization_transcription_gt_v2.zip
    

  • Step2 数据集规整
  1. Gif 文件转png

     import os
     from PIL import Image
     
     git_filename = 'img_401.gif'
     # 使用Image模块的open()方法打开gif动态图像时，默认是第一帧
     im = Image.open(git_filename)
     try:
         while True:
             # 保存当前帧图片
             current = im.tell()
             im.save('img_401.png')
             # 获取下一帧图片
             # im.seek(current+1)
     except EOFError:
         pass
     

  2. 按前言的目录组织，将下载的文件拷贝到对应文件目录

  • Step3 产生数据索引文件train.txt、test.txt

    from __future__ import absolute_import
    from __future__ import division
    from __future__ import print_function
    from __future__ import unicode_literals
    
    import os, sys
    import numpy as np
    
    icdar2017_root_dir = os.getcwd()
    icdar2017_train_img_dir = os.path.join(icdar2017_root_dir, 'train', 'img')
    icdar2017_train_gt_dir = os.path.join(icdar2017_root_dir, 'train', 'gt')
    
    icdar2017_test_img_dir = os.path.join(icdar2017_root_dir, 'test', 'img')
    icdar2017_test_gt_dir = os.path.join(icdar2017_root_dir, 'test', 'gt')
    
    print(f'icdar2017_root_dir:{icdar2017_root_dir}')
    print(f'icdar2017_train_img_dir:{icdar2017_train_img_dir}')
    print(f'icdar2017_train_gt_dir:{icdar2017_train_gt_dir}')
    
    print(f'icdar2017_test_img_dir:{icdar2017_test_img_dir}')
    print(f'icdar2017_test_gt_dir:{icdar2017_test_gt_dir}')
    print('*'*80)
    
    with open('train.txt', 'w') as f:
        imgs = os.listdir(icdar2017_train_img_dir)
        print(f'write train images:{len(imgs)}')
        for img in imgs:        
            img_path = os.path.join(icdar2017_train_img_dir, img)
            if img_path.endswith('jpg'):
                gt_name = 'gt_' + img.replace('jpg', 'txt')
            elif img_path.endswith('png'):
                gt_name = 'gt_' + img.replace('png', 'txt')
    
            gt_path = os.path.join(icdar2017_train_gt_dir, gt_name)
            f.write(img_path + '\t' + gt_path + '\n')
    
    with open('test.txt', 'w') as f:
        imgs = os.listdir(icdar2017_test_img_dir)
        print(f'write test images:{len(imgs)}')
        for img in imgs:
            img_path = os.path.join(icdar2017_test_img_dir, img)
            if img.endswith('jpg'):
                gt_name = 'gt_' + img.replace('jpg', 'txt')
            elif img.endswith('png'):
                gt_name = 'gt_' + img.replace('png', 'txt')
                
            gt_path = os.path.join(icdar2017_test_gt_dir, gt_name)
            f.write(img_path + '\t' + gt_path + '\n')        
            
    print(f'****************** generate label list done ******************')  
    

  • Step4 测试
    测试代码如ICDAR2015

  • ICDAR2019

    • Step1 数据集下载

      wget http://datasets.cvc.uab.es/rrc/ImagesPart1.zip
      wget http://datasets.cvc.uab.es/rrc/ImagesPart2.zip
      wget http://datasets.cvc.uab.es/rrc/train_gt_t13.zip
      wget http://datasets.cvc.uab.es/rrc/MLT19_TestImagesPart1.zip
      wget http://datasets.cvc.uab.es/rrc/MLT19_TestImagesPart2.zip
      

      Note that this task only requires localization results (as indicated in results format in the tasks page), but the ground truth also provides the script id of each bounding box and the transcription. This extra information will be needed in Tasks 3 and 4.   
      Extra information about the training set (may be useful for researchers who focus on one or only few languages, not all of the multi-lingual set):   	
      The 10,000 images are ordered in the training set such that: each consecutive 1000 images contain text of one main language (and it may of course contain additional text from 1 or 2 other languages, all from the set of the 10 languages)    
      - 00001 - 01000:  Arabic
      - 01001 - 02000:  English
      - 02001 - 03000:  French
      - 03001 - 04000:  Chinese
      - 04001 - 05000:  German
      - 05001 - 06000:  Korean
      - 06001 - 07000:  Japanese
      - 07001 - 08000:  Italian
      - 08001 - 09000:  Bangla
      - 09001 - 10000:  Hindi
      

    • Step2 数据集规整
    • Step3 产生数据索引文件train.txt、test.txt
    • Step4 测试
      测试代码如ICDAR2015

  • ICDAR2019

    • Step1 数据集下载
    • Step2 数据集规整
    • Step3 产生数据索引文件train.txt、test.txt
    • Step4 测试

  • ICDAR2019

    • Step1 数据集下载
    • Step2 数据集规整
    • Step3 产生数据索引文件train.txt、test.txt
    • Step4 测试

  • ICDAR2019

    • Step1 数据集下载
    • Step2 数据集规整
    • Step3 产生数据索引文件train.txt、test.txt
    • Step4 测试