icdar2015 文本定位数据集

Mr.Q

已于 2023-03-02 09:37:54 修改

阅读量9.4k

点赞数 5

分类专栏： OCR 文章标签： OCR

于 2022-04-13 18:42:14 首次发布

本文链接：https://blog.csdn.net/jizhidexiaoming/article/details/124149164

版权

OCR 专栏收录该内容

12 篇文章 0 订阅

订阅专栏

2.2 instances_training.json

3. 生成json脚本

3.1 载入图片和标注信息

3.2 保存成json

1. 下载

1.1 官网链接

目前官网有4个任务，第1个任务就是文本定位：

Downloads - Incidental Scene Text - Robust Reading Competitionhttps://rrc.cvc.uab.es/?ch=4&com=downloads官网下载下来如下：

1.2 网盘链接

将官网的文件处理成能够训练的格式：

mkdir icdar2015 && cd icdar2015
mkdir imgs && mkdir annotations
# For images,
mv ch4_training_images imgs/training
mv ch4_test_images imgs/test
# For annotations,
mv ch4_training_localization_transcription_gt annotations/training
mv Challenge4_Test_Task1_GT annotations/test

python tools/data/textdet/icdar_converter.py /path/to/icdar2015 -o /path/to/icdar2015 -d icdar2015 --split-list training test

链接：https://pan.baidu.com/s/1yypel9p30ws6mWYZd14-iw
提取码：inc5
--来自百度网盘超级会员V4的分享

文件格式如下：

annotations
    test
    training
imgs
    test
    training
instances_test.json
instances_training.json

2. 文件内容介绍

imgs文件夹就是一些图片，如下。

annotations文件夹是标注文件。

2.1 gt_img_1.txt

377,117,463,117,465,130,378,130,Genaxis Theatre
493,115,519,115,519,131,493,131,[06]
374,155,409,155,409,170,374,170,###
492,151,551,151,551,170,492,170,62-03
376,198,422,198,422,212,376,212,Carpark
494,190,539,189,539,205,494,206,###
374,1,494,0,492,85,372,86,###

对应的图片：

（1）377,117,463,117,465,130,378,130,Genaxis Theatre

是上面图片上的文本标注（多边形4个顶点+文本内容）：x1,y1,x2,y2,x3,y3,x4,y4，文本内容。

（2）###

#号是忽略标注。将会丢弃，不参与训练。

代码读取gt.txt文件，并画出多边形

import cv2
import numpy

if __name__ == '__main__':
    img_path = r"E:\img_1.jpg"
    gt_path = r"E:\gt_img_1.txt"

    img = cv2.imread(img_path)
    with open(gt_path, mode="r", encoding='utf-8-sig') as f:
        gt_list = f.readlines()
        for gt in gt_list:
            gt = gt.split(',')
            del gt[-1]
            print(gt[0])
            pt_list = []
            for i in range(0, len(gt)-1, 2):
                pt_list.append([int(gt[i]), int(gt[i+1])])
            pt_array = numpy.array(pt_list)
            cv2.drawContours(img, [pt_array], -1, (0, 0, 255), 3)
    f.close()
    cv2.namedWindow("img", cv2.WINDOW_NORMAL), cv2.imshow("img", img), cv2.waitKey()

2.2 instances_training.json

{
"images": [
    # images里面存放的是字典，每个字典内容：图片路径、高、宽、对应的字符分割标注文件路径、图片id；
	{"file_name": "training/img_927.jpg", "height": 720, "width": 1280, "segm_file": "training/gt_img_927.txt", "id": 0}, 
	{"file_name": "training/img_573.jpg", "height": 720, "width": 1280, "segm_file": "training/gt_img_573.txt", "id": 1}, 
	{"file_name": "training/img_510.jpg", "height": 720, "width": 1280, "segm_file": "training/gt_img_510.txt", "id": 2}, 
	{"file_name": "training/img_468.jpg", "height": 720, "width": 1280, "segm_file": "training/gt_img_468.txt", "id": 3}, 
	{"file_name": "training/img_827.jpg", "height": 720, "width": 1280, "segm_file": "training/gt_img_827.txt", "id": 4}
    ], 

"categories": [{"id": 1, "name": "text"}],   # 这个数据集只有一个类别，即文本，id设置为1.

"annotations": [
	{"iscrowd": 1, 
	"category_id": 1, 
	"bbox": [216.0, 222.0, 93.0, 37.0], 
	"area": 2054.0, 
	"segmentation": [[217, 235, 304, 222, 309, 243, 216, 259]], 
	"image_id": 0,
	"id": 0
    }, 
    
    # 所有图片的所有标注信息
    # "iscrowd"：等于1则是无用标注；默认等于0是有用标注；
    # "category": 默认等于1；
    # "bbox"：文本区域最小外接矩形(min_x, min_y, w, h)；
    # "area": 多边形文本区域面积；
    # "segmentation": 多边形文本区域的4个顶点坐标：[[x1, y1], [x2, y2], ...]
    # "image_id": 此条标注所在的图片id；
    # "id"：此条标注id.

    {"iscrowd": 1, "category_id": 1, "bbox": [296.0, 417.0, 85.0, 45.0], "area": 2303.0, "segmentation": [[296, 434, 374, 417, 381, 445, 303, 462]], "image_id": 0, "id": 1}, 
    {"iscrowd": 0, "category_id": 1, "bbox": [374.0, 163.0, 27.0, 14.0], "area": 314.0, "segmentation": [[374, 165, 400, 163, 401, 175, 375, 177]], "image_id": 0, "id": 2}, 
    {"iscrowd": 0, "category_id": 1, "bbox": [400.0, 161.0, 31.0, 13.0], "area": 301.5, "segmentation": [[400, 163, 428, 161, 431, 171, 403, 174]], "image_id": 0, "id": 3}, 
]

}

3. 生成json脚本

3.1 载入图片和标注信息

def load_img_info(files, dataset):
    """Load the information of one image.

    Args:
        files(tuple): The tuple of (img_file, groundtruth_file)
        dataset(str): Dataset name, icdar2015 or icdar2017

    Returns:
        img_info(dict): The dict of the img and annotation information
    """
    assert isinstance(files, tuple)
    assert isinstance(dataset, str)
    assert dataset

    img_file, gt_file = files
    # read imgs with ignoring orientations
    img = mmcv.imread(img_file, 'unchanged')

    if dataset == 'icdar2017':
        gt_list = list_from_file(gt_file)
    elif dataset == 'icdar2015':
        gt_list = list_from_file(gt_file, encoding='utf-8-sig')  # txt2list
    else:
        raise NotImplementedError(f'Not support {dataset}')

    anno_info = []
    for line in gt_list:
        # each line has one ploygen (4 vetices), and others.
        # e.g., 695,885,866,888,867,1146,696,1143,Latin,9
        line = line.strip()  # "x1, y1, x2, y2, ..., Text"
        strs = line.split(',')
        category_id = 1
        xy = [int(x) for x in strs[0:8]]  # 前面八个数字是文本位置坐标。str2int: [x1, y1, x2, y2, ...]
        coordinates = np.array(xy).reshape(-1, 2)  # list.size=8 转 numpy.size=(4,2): [[x1, y1], [x2, y2], ...]
        polygon = Polygon(coordinates)
        iscrowd = 0  # 遍历一张图片上的所有标注，iscrowd等于0，说明是有用标注。
        # set iscrowd to 1 to ignore 1.
        if (dataset == 'icdar2015'
                and strs[8] == '###') or (dataset == 'icdar2017'
                                          and strs[9] == '###'):
            iscrowd = 1  # 等于1，说明是无用标注。
            print('ignore text')

        area = polygon.area
        # convert to COCO style XYWH format
        min_x, min_y, max_x, max_y = polygon.bounds  # minimum bounding region
        bbox = [min_x, min_y, max_x - min_x, max_y - min_y]

        anno = dict(
            iscrowd=iscrowd,  # 1
            category_id=category_id,  # 1
            bbox=bbox,  # 文本最小外接矩形。
            area=area,  # 多边形文本区域面积
            segmentation=[xy])  # 多边形文本区域顶点
        anno_info.append(anno)
    split_name = osp.basename(osp.dirname(img_file))
    img_info = dict(
        # remove img_prefix for filename
        file_name=osp.join(split_name, osp.basename(img_file)),
        height=img.shape[0],
        width=img.shape[1],
        anno_info=anno_info,
        segm_file=osp.join(split_name, osp.basename(gt_file)))
    return img_info

传入图片路径和对应的标注.txt路径，返回字典

3.2 保存成json

将前面得到的img_info字典信息，保存成json文件。

def convert_annotations(image_infos, out_json_name):
    """Convert the annotation into coco style.

    Args:
        image_infos(list): The list of image information dicts
        out_json_name(str): The output json filename

    Returns:
        out_json(dict): The coco style dict
    """
    assert isinstance(image_infos, list)
    assert isinstance(out_json_name, str)
    assert out_json_name  # json文件完整路径

    out_json = dict()
    img_id = 0
    ann_id = 0
    out_json['images'] = []
    out_json['categories'] = []
    out_json['annotations'] = []
    for image_info in image_infos:  # 遍历所有图片的信息
        image_info['id'] = img_id  # image_info添加img_id信息
        anno_infos = image_info.pop('anno_info')
        out_json['images'].append(image_info)
        for anno_info in anno_infos:  # anno_info添加image_id、id信息
            anno_info['image_id'] = img_id  
            anno_info['id'] = ann_id
            out_json['annotations'].append(anno_info)
            ann_id += 1
        img_id += 1
    
    # 下面两行是额外的信息，暂时不知用处。
    cat = dict(id=1, name='text')
    out_json['categories'].append(cat)

    if len(out_json['annotations']) == 0:  # 所有图片都没有文本标注，则删除annotations key.
        out_json.pop('annotations')
    mmcv.dump(out_json, out_json_name)

    return out_json