VOC数据集生成代码使用说明

#split.py 文件 输入格式为images ,和标签txt文件，txt中的数据为坐标值共8个。

import os
import numpy as np
import math
import cv2 as cv
import imageio

#path = '/media/D/code/OCR/text-detection-ctpn/data/mlt_english+chinese/image'
#path = '/home/chendali1/Gsj/text-detection-ctpn-master/prepare_training_data/image/image_1000/'
path='/home/chendali1/Gsj/prepare_training_data/ICDAR/images_train/'
#gt_path = '/home/chendali1/Gsj/text-detection-ctpn-master/prepare_training_data/label/labelDigit1000/'
gt_path='/home/chendali1/Gsj/prepare_training_data/ICDAR/result_train/'
out_path = 're_image'
if not os.path.exists(out_path):
    os.makedirs(out_path)
files = os.listdir(path)
files.sort()
#files=files[:100]
for file in files:
    _, basename = os.path.split(file)
    if basename.lower().split('.')[-1] not in ['jpg', 'png']:
        continue
    stem, ext = os.path.splitext(basename)
    
    
    #stem=stem0.split('_')[2]
    
    gt_file = os.path.join(gt_path, stem+'.txt')
    img_path = os.path.join(path, file)
    print(img_path)
    #print(gt_file)
    img = cv.imread(img_path)
    if img is None:
        print('****************************')
        print('Image ' + img_path + ' may be a bad picture!')
        print('****************************')
        newname = os.path.join(path,stem+'.gif')
        os.rename(img_path,newname)
        img_path=newname
        print(img_path)
        print('Try read with imageio.')
        gif = imageio.mimread(img_path)
        if gif is None:
            print('****************************')
            print("Image " + img_path + " can't be read!")
            print('****************************')
            
        print('Read success!')
        img = cv.cvtColor(gif[0], cv.COLOR_RGB2BGR)
    
    img_size = img.shape
    im_size_min = np.min(img_size[0:2])
    im_size_max = np.max(img_size[0:2])

    im_scale = float(600) / float(im_size_min)
    if np.round(im_scale * im_size_max) > 1200:
        im_scale = float(1200) / float(im_size_max)
    re_im = cv.resize(img, None, None, fx=im_scale, fy=im_scale, interpolation=cv.INTER_LINEAR)
    re_size = re_im.shape
    cv.imwrite(os.path.join(out_path, stem) + '.jpg', re_im)

    with open(gt_file, 'r') as f:
        lines = f.readlines()
    for line in lines:
        splitted_line = line.strip().lower().split(',')
        pt_x = np.zeros((4, 1))
        pt_y = np.zeros((4, 1))
        pt_x[0, 0] = int(float(splitted_line[0]) / img_size[1] * re_size[1])
        pt_y[0, 0] = int(float(splitted_line[1]) / img_size[0] * re_size[0])
        pt_x[1, 0] = int(float(splitted_line[2]) / img_size[1] * re_size[1])
        pt_y[1, 0] = int(float(splitted_line[3]) / img_size[0] * re_size[0])
        pt_x[2, 0] = int(float(splitted_line[4]) / img_size[1] * re_size[1])
        pt_y[2, 0] = int(float(splitted_line[5]) / img_size[0] * re_size[0])
        pt_x[3, 0] = int(float(splitted_line[6]) / img_size[1] * re_size[1])
        pt_y[3, 0] = int(float(splitted_line[7]) / img_size[0] * re_size[0])

        ind_x = np.argsort(pt_x, axis=0)
        pt_x = pt_x[ind_x]
        pt_y = pt_y[ind_x]

        if pt_y[0] < pt_y[1]:
            pt1 = (pt_x[0], pt_y[0])
            pt3 = (pt_x[1], pt_y[1])
        else:
            pt1 = (pt_x[1], pt_y[1])
            pt3 = (pt_x[0], pt_y[0])

        if pt_y[2] < pt_y[3]:
            pt2 = (pt_x[2], pt_y[2])
            pt4 = (pt_x[3], pt_y[3])
        else:
            pt2 = (pt_x[3], pt_y[3])
            pt4 = (pt_x[2], pt_y[2])

        xmin = int(min(pt1[0], pt2[0]))
        ymin = int(min(pt1[1], pt2[1]))
        xmax = int(max(pt2[0], pt4[0]))
        ymax = int(max(pt3[1], pt4[1]))

        if xmin < 0:
            xmin = 0
        if xmax > re_size[1] - 1:
            xmax = re_size[1] - 1
        if ymin < 0:
            ymin = 0
        if ymax > re_size[0] - 1:
            ymax = re_size[0] - 1

        width = xmax - xmin
        height = ymax - ymin

        # reimplement
        step = 16.0
        x_left = []
        x_right = []
        x_left.append(xmin)
        x_left_start = int(math.ceil(xmin / 16.0) * 16.0)
        if x_left_start == xmin:
            x_left_start = xmin + 16
        for i in np.arange(x_left_start, xmax, 16):
            x_left.append(i)
        x_left = np.array(x_left)

        x_right.append(x_left_start - 1)
        for i in range(1, len(x_left) - 1):
            x_right.append(x_left[i] + 15)
        x_right.append(xmax)
        x_right = np.array(x_right)

        idx = np.where(x_left == x_right)
        x_left = np.delete(x_left, idx, axis=0)
        x_right = np.delete(x_right, idx, axis=0)

        if not os.path.exists('label_tmp'):
            os.makedirs('label_tmp')
        with open(os.path.join('label_tmp', stem) + '.txt', 'a') as f:
            #for i in range(len(x_left)):
            f.writelines("tianchi	")
            f.writelines(str(int( pt_x[0, 0])))
            f.writelines("	")
            f.writelines(str(int( pt_y[0, 0])))
            f.writelines("	")
            f.writelines(str(int( pt_x[1, 0])))
            f.writelines("	")
            f.writelines(str(int( pt_y[1, 0])))
            f.writelines("	")
            f.writelines(str(int( pt_x[2, 0])))
            f.writelines("	")
            f.writelines(str(int( pt_y[2, 0])))
            f.writelines("	")
            f.writelines(str(int( pt_x[3, 0])))
            f.writelines("	")
            f.writelines(str(int( pt_y[3, 0])))                
            f.writelines("
")

#ToVoc.py 上述执行完后直接运行这个脚本文件完美生成VOC文件
from xml.dom.minidom import Document
import cv2
import os
import glob
import shutil
import numpy as np

def generate_xml(name, lines, img_size, class_sets, doncateothers=True):
    doc = Document()

    def append_xml_node_attr(child, parent=None, text=None):
        ele = doc.createElement(child)
        if not text is None:
            text_node = doc.createTextNode(text)
            ele.appendChild(text_node)
        parent = doc if parent is None else parent
        parent.appendChild(ele)
        return ele

    img_name = name + '.jpg'
    # create header
    annotation = append_xml_node_attr('annotation')
    append_xml_node_attr('folder', parent=annotation, text='tianchi')
    append_xml_node_attr('filename', parent=annotation, text=img_name)
    source = append_xml_node_attr('source', parent=annotation)
    append_xml_node_attr('database', parent=source, text='coco_text_database')
    append_xml_node_attr('annotation', parent=source, text='tianchi')
    append_xml_node_attr('image', parent=source, text='tianchi')
    append_xml_node_attr('flickrid', parent=source, text='000000')
    owner = append_xml_node_attr('owner', parent=annotation)
    append_xml_node_attr('name', parent=owner, text='ms')
    size = append_xml_node_attr('size', annotation)
    append_xml_node_attr('width', size, str(img_size[1]))
    append_xml_node_attr('height', size, str(img_size[0]))
    append_xml_node_attr('depth', size, str(img_size[2]))
    append_xml_node_attr('segmented', parent=annotation, text='0')

    # create objects
    objs = []
    for line in lines:
        splitted_line = line.strip().lower().split()
        cls = splitted_line[0].lower()
        if not doncateothers and cls not in class_sets:
            continue
        cls = 'dontcare' if cls not in class_sets else cls
        if cls == 'dontcare':
            continue
        obj = append_xml_node_attr('object', parent=annotation)
        occlusion = int(0)
        x1, y1, x2, y2 = int(float(splitted_line[1]) + 1), int(float(splitted_line[2]) + 1), 
                         int(float(splitted_line[3]) + 1), int(float(splitted_line[4]) + 1)
        x0,y0,x1,y1,x2,y2,x3,y3 = int(float(splitted_line[1])+1),int(float(splitted_line[2])+1),
        int(float(splitted_line[3])+1),int(float(splitted_line[4])+1),int(float(splitted_line[5])+1),
        int(float(splitted_line[6])+1),int(float(splitted_line[7])+1),int(float(splitted_line[8])+1)
        truncation = float(0)
        difficult = 1 if _is_hard(cls, truncation, occlusion, x1, y1, x2, y2) else 0
        truncted = 0 if truncation < 0.5 else 1

        append_xml_node_attr('name', parent=obj, text=cls)
        append_xml_node_attr('pose', parent=obj, text='none')
        append_xml_node_attr('truncated', parent=obj, text=str(truncted))
        append_xml_node_attr('difficult', parent=obj, text=str(int(difficult)))
        bb = append_xml_node_attr('bndbox', parent=obj)
        append_xml_node_attr('x0', parent=bb, text=str(int(x0)))
        append_xml_node_attr('y0', parent=bb, text=str(y0))
        append_xml_node_attr('x1', parent=bb, text=str(x1))
        append_xml_node_attr('y1', parent=bb, text=str(y1))
        append_xml_node_attr('x1', parent=bb, text=str(x2))
        append_xml_node_attr('y1', parent=bb, text=str(y2))
        append_xml_node_attr('x1', parent=bb, text=str(x3))
        append_xml_node_attr('y1', parent=bb, text=str(y3))

        o = {'class': cls, 'box': np.asarray([x0, y0,x1,y1, x2, y2,x3,y3], dtype=float), 
             'truncation': truncation, 'difficult': difficult, 'occlusion': occlusion}
        objs.append(o)

    return doc, objs


def _is_hard(cls, truncation, occlusion, x1, y1, x2, y2):
    hard = False
    if y2 - y1 < 25 and occlusion >= 2:
        hard = True
        return hard
    if occlusion >= 3:
        hard = True
        return hard
    if truncation > 0.8:
        hard = True
        return hard
    return hard


def build_voc_dirs(outdir):
    mkdir = lambda dir: os.makedirs(dir) if not os.path.exists(dir) else None
    mkdir(outdir)
    mkdir(os.path.join(outdir, 'Annotations'))
    mkdir(os.path.join(outdir, 'ImageSets'))
    mkdir(os.path.join(outdir, 'ImageSets', 'Layout'))
    mkdir(os.path.join(outdir, 'ImageSets', 'Main'))
    mkdir(os.path.join(outdir, 'ImageSets', 'Segmentation'))
    mkdir(os.path.join(outdir, 'JPEGImages'))
    mkdir(os.path.join(outdir, 'SegmentationClass'))
    mkdir(os.path.join(outdir, 'SegmentationObject'))
    return os.path.join(outdir, 'Annotations'), os.path.join(outdir, 'JPEGImages'), os.path.join(outdir, 'ImageSets',
                                                                                                 'Main')


if __name__ == '__main__':
    _outdir = 'TEXTVOC/VOC2007'
    _draw = bool(0)
    _dest_label_dir, _dest_img_dir, _dest_set_dir = build_voc_dirs(_outdir)
    _doncateothers = bool(1)
    for dset in ['train']:
        _labeldir = 'label_tmp'
        _imagedir = 're_image'
        class_sets = ('tianchi', 'dontcare')
        class_sets_dict = dict((k, i) for i, k in enumerate(class_sets))
        allclasses = {}
        fs = [open(os.path.join(_dest_set_dir, cls + '_' + dset + '.txt'), 'w') for cls in class_sets]
        ftrain = open(os.path.join(_dest_set_dir, dset + '.txt'), 'w')

        files = glob.glob(os.path.join(_labeldir, '*.txt'))
        files.sort()
        for file in files:
            path, basename = os.path.split(file)
            stem, ext = os.path.splitext(basename)
            with open(file, 'r') as f:
                lines = f.readlines()
            img_file = os.path.join(_imagedir, stem + '.jpg')

            print(img_file)
            img = cv2.imread(img_file)
            img_size = img.shape

            doc, objs = generate_xml(stem, lines, img_size, class_sets=class_sets, doncateothers=_doncateothers)

            cv2.imwrite(os.path.join(_dest_img_dir, stem + '.jpg'), img)
            xmlfile = os.path.join(_dest_label_dir, stem + '.xml')
            with open(xmlfile, 'w') as f:
                f.write(doc.toprettyxml(indent='    '))

            ftrain.writelines(stem + '
')

            cls_in_image = set([o['class'] for o in objs])

            for obj in objs:
                cls = obj['class']
                allclasses[cls] = 0 
                    if not cls in list(allclasses.keys()) else allclasses[cls] + 1

            for cls in cls_in_image:
                if cls in class_sets:
                    fs[class_sets_dict[cls]].writelines(stem + ' 1
')
            for cls in class_sets:
                if cls not in cls_in_image:
                    fs[class_sets_dict[cls]].writelines(stem + ' -1
')


        (f.close() for f in fs)
        ftrain.close()

        print('~~~~~~~~~~~~~~~~~~~')
        print(allclasses)
        print('~~~~~~~~~~~~~~~~~~~')
        shutil.copyfile(os.path.join(_dest_set_dir, 'train.txt'), os.path.join(_dest_set_dir, 'val.txt'))
        shutil.copyfile(os.path.join(_dest_set_dir, 'train.txt'), os.path.join(_dest_set_dir, 'trainval.txt'))
        for cls in class_sets:
            shutil.copyfile(os.path.join(_dest_set_dir, cls + '_train.txt'),
                            os.path.join(_dest_set_dir, cls + '_trainval.txt'))
            shutil.copyfile(os.path.join(_dest_set_dir, cls + '_train.txt'),
                            os.path.join(_dest_set_dir, cls + '_val.txt'))