数据集转换为VOC格式

之前跑yolov5时，找到的将数据集转换为VOC格式的程序，这里做个记录供之后使用。

# coding:utf-8

import os
import xml.etree.ElementTree as ET
from os import getcwd
import random
import argparse
import cv2
import numpy as np
import shutil

path = '/home/dataset/Data'
if path == '':
    path = os.getcwd()    # 如果数据集就在给目录下就直接用这个

sets = ['train', 'val', 'test']
classes = []  # 改为自己的类别
trainval_percent = 0.5  # 训练集和验证集所占比例。 这里没有划分测试集
train_percent = 0.8  # 训练集所占比例，可自己进行调整

parser = argparse.ArgumentParser()
# xml/txt文件的地址，根据自己的数据进行修改 xml/txt一般存放在Annotations/labels下
parser.add_argument('--xml_path', default=os.path.join(path, 'Annotations'), type=str,
                    help='input txt label path')
parser.add_argument('--xml2txt_path', default=os.path.join(path, 'labels'), type=str,
                    help='input txt label path')
# 数据集的划分，地址选择自己数据下的ImageSets/Main
parser.add_argument('--txt_path', default=os.path.join(path, 'ImageSets/Main'), type=str,
                    help='output txt label path')
parser.add_argument('--img_path', default=os.path.join(path, 'JPEGImages'), type=str,
                    help='input img path')
parser.add_argument('--img2_path', default=os.path.join(path, 'images'), type=str,
                    help='output img path')
parser.add_argument('--img_size', default=640, type=int, help='img size(pixels)')
opt = parser.parse_args()


xmlfilepath = opt.xml_path
if not os.path.exists(xmlfilepath):
    os.mkdir(xmlfilepath)
xml2txtpath = opt.xml2txt_path
if not os.path.exists(xml2txtpath):
    os.mkdir(xml2txtpath)
if not os.path.exists(opt.txt_path):
    os.mkdir(opt.txt_path)
txtsavepath = os.path.join(opt.txt_path)
imgfilepath = opt.img_path
if not os.path.exists(imgfilepath):
    os.mkdir(imgfilepath)
imgfile2path = opt.img2_path
if not os.path.exists(imgfile2path):
    os.mkdir(imgfile2path)
total_xml = os.listdir(xmlfilepath)


def create_txt():
    num = len(total_xml)
    list_index = range(num)
    tv = int(num * trainval_percent)
    tr = int(tv * train_percent)
    trainval = random.sample(list_index, tv)
    train = random.sample(trainval, tr)

    file_trainval = open(txtsavepath + '/trainval.txt', 'w')
    file_test = open(txtsavepath + '/test.txt', 'w')
    file_train = open(txtsavepath + '/train.txt', 'w')
    file_val = open(txtsavepath + '/val.txt', 'w')

    img_train = open(path + '/train.txt', 'w')
    img_val = open(path + '/val.txt', 'w')
    img_test = open(path + '/test.txt', 'w')


    for i in list_index:
        name = total_xml[i][:-4] + '\n'
        img_path = os.path.join(opt.img2_path, total_xml[i][:-4] + '.jpg' + '\n')
        if i in trainval:
            file_trainval.write(name)
            if i in train:
                file_train.write(name)
                if os.path.exists(img_path.split('\n')[0]):
                    img_train.write(img_path)
                else:
                    img_path = os.path.join(opt.img2_path, total_xml[i][:-4] + '.png' + '\n')
                    img_train.write(img_path)
            else:
                file_val.write(name)
                if os.path.exists(img_path.split('\n')[0]):
                    img_val.write(img_path)
                else:
                    img_path = os.path.join(opt.img2_path, total_xml[i][:-4] + '.png' + '\n')
                    img_val.write(img_path)
        else:
            file_test.write(name)
            if os.path.exists(img_path.split('\n')[0]):
                img_test.write(img_path)
            else:
                img_path = os.path.join(opt.img2_path, total_xml[i][:-4] + '.png' + '\n')
                img_test.write(img_path)
    file_trainval.close()
    file_train.close()
    file_val.close()
    file_test.close()
    img_train.close()
    img_val.close()
    img_test.close()


def img_augument(img_dir, save_img_dir, labels_dir):
        if os.path.exists(img_dir):
            img_name= img_dir.split('/')[-1]
        else:
            img_name = img_dir.split('.')[0] + ".png"
        print("process %s" % os.path.join(img_dir))
        img = cv2.imread(img_dir)
        h, w = img.shape[0:2]
        input_size = opt.img_size
        # 因为数据集内原始图像的尺寸是不定的，所以需要进行适当的padding，将原始图像padding成宽高一致的正方形
        # 然后再将Padding后的正方形图像缩放成448x448
        padw, padh = 0, 0  # 要记录宽高方向的padding具体数值，因为padding之后需要调整bbox的位置信息
        if h > w:
            padw = (h - w) // 2
            img = np.pad(img, ((0, 0), (padw, padw), (0, 0)), 'constant', constant_values=0)
        elif w > h:
            padh = (w - h) // 2
            img = np.pad(img, ((padh, padh), (0, 0), (0, 0)), 'constant', constant_values=0)
        img = cv2.resize(img, (input_size, input_size))
        cv2.imwrite(save_img_dir, img)
        # 读取图像对应的bbox信息，按1维的方式储存，每5个元素表示一个bbox的(cls,xc,yc,w,h)
        with open(os.path.join(xml2txtpath, img_name.split('.')[0] + ".txt"), 'r') as f:
            bbox = f.read().split('\n')
        bbox = [x.split() for x in bbox]
        bbox = [float(x) for y in bbox for x in y]
        if len(bbox) % 5 != 0:
            raise ValueError("File:"
                             + os.path.join(labels_dir, img_name.split('.')[0] + ".txt") + "——bbox Extraction Error!")

        # 根据padding、图像增广等操作，将原始的bbox数据转换为修改后图像的bbox数据
        if padw != 0:
            for i in range(len(bbox) // 5):
                bbox[i * 5 + 1] = (bbox[i * 5 + 1] * w + padw) / h
                bbox[i * 5 + 3] = (bbox[i * 5 + 3] * w) / h
                cv2.rectangle(img, (int(bbox[1] * input_size - bbox[3] * input_size / 2),
                                    int(bbox[2] * input_size - bbox[4] * input_size / 2)),
                              (int(bbox[1] * input_size + bbox[3] * input_size / 2),
                               int(bbox[2] * input_size + bbox[4] * input_size / 2)), (0, 0, 255))
        elif padh != 0:
            for i in range(len(bbox) // 5):
                bbox[i * 5 + 2] = (bbox[i * 5 + 2] * h + padh) / w
                bbox[i * 5 + 4] = (bbox[i * 5 + 4] * h) / w
                cv2.rectangle(img, (int(bbox[1] * input_size - bbox[3] * input_size / 2),
                                    int(bbox[2] * input_size - bbox[4] * input_size / 2)),
                              (int(bbox[1] * input_size + bbox[3] * input_size / 2),
                               int(bbox[2] * input_size + bbox[4] * input_size / 2)), (0, 0, 255))
        # 此处可以写代码验证一下，查看padding后修改的bbox数值是否正确，在原图中画出bbox检验
        # cv2.imshow("bbox-%d" % int(bbox[0]), img)
        # cv2.waitKey(0)
        with open(os.path.join(labels_dir, img_name.split('.')[0] + ".txt"), 'w') as f:
            for i in range(len(bbox) // 5):
                bbox = [str(x) for x in bbox[i * 5:(i * 5 + 5)]]
                str_context = " ".join(bbox) + '\n'
                f.write(str_context)


def rename(file, name):
    imgpath = os.path.join(imgfilepath, name[:-1])
    if os.path.exists(imgpath + '.jpg'):
        imgpath = imgpath + '.jpg'
        name_2 = os.path.join(imgfile2path, name[:-1] + '.jpg')
    elif os.path.exists(name + '.png'):
        imgpath = imgpath + '.png'
        name_2 = os.path.join(imgfile2path, name[:-1] + '.png')
    else:
        name_2 = None
        print("cannot find picture %s" % imgpath)
    return imgpath, name_2


def copy_img():
    for x in os.listdir(txtsavepath):
        with open(txtsavepath + os.sep + x, 'r') as txtread:
            for name in txtread.readlines():
                name_1, name_2 = rename(x, name)
                if name_2 is None:
                    continue
                if not os.path.exists(name_2):
                    img_augument(name_1, name_2, xmlfilepath)
                    shutil.copy(name_1, name_2)
                    print("picture%s is finish" % name_2)
                else:
                    continue


def convert(size, box):
    dw = 1. / (size[0])
    dh = 1. / (size[1])
    x = (box[0] + box[1]) / 2.0 - 1
    y = (box[2] + box[3]) / 2.0 - 1
    w = box[1] - box[0]
    h = box[3] - box[2]
    x = x * dw
    w = w * dw
    y = y * dh
    h = h * dh
    return x, y, w, h


def convert_annotation(path, image_id):
    in_file = open(os.path.join(path, 'Annotations/%s.xml' % (image_id.split('.')[0])), encoding='UTF-8')
    out_file = open(os.path.join(path, 'labels/%s.txt' % (image_id.split('.')[0])), 'w')
    tree = ET.parse(in_file)
    root = tree.getroot()
    size = root.find('size')
    w = int(size.find('width').text)
    h = int(size.find('height').text)
    for obj in root.iter('object'):
        difficult = obj.find('difficult').text
        # difficult = obj.find('Difficult').text
        cls = obj.find('name').text
        if cls not in classes or int(difficult) == 1:
            continue
        cls_id = classes.index(cls)
        xmlbox = obj.find('bndbox')
        b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text),
             float(xmlbox.find('ymax').text))
        b1, b2, b3, b4 = b
        # 标注越界修正
        if b2 > w:
            b2 = w
        if b4 > h:
            b4 = h
        b = (b1, b2, b3, b4)
        bb = convert((w, h), b)
        out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')


def make_label_txt():
    """在labels文件夹下创建image_id.txt，对应每个image_id.xml提取出的bbox信息"""
    filenames = os.listdir(os.path.join(path, 'Annotations'))
    for file in filenames:
        convert_annotation(path, file)


def xml2yolotxt():
    for image_set in sets:
        # 这里是绝对路径，需要根据自己的情况修改
        if not os.path.exists(os.path.join(path, 'labels/')):
            os.makedirs(os.path.join(path, 'labels/'))
        image_ids = open(os.path.join(path, 'ImageSets/Main/%s.txt' % (image_set))).read().strip().split()

        if not os.path.exists(os.path.join(path, 'images/')):
            os.makedirs(os.path.join(path, 'images/'))

        list_file = open(os.path.join(path, '%s.txt' % (image_set)), 'w')
        for image_id in image_ids:
            list_file.write(os.path.join(path, 'images/%s.jpg\n' % (image_id.split('/')[-1])))
            convert_annotation(path, image_id.split('/')[-1])
        list_file.close()


make_label_txt()
create_txt()
copy_img()
xml2yolotxt()