【工具包使用】xml文件解析

前言

 

数据集简介

aistudio.baidu.com

这个数据集是从网络上爬取的，图像尺寸相差较大，小的80*90，大的3000*4300，异质性较大，解析之后训练效果非常差；

 

数据集解析

# 20231019: driver behavior.
# classid: 0-phone, 1-smoke, 2-drink.
import csv
import xml.etree.ElementTree as ET
import cv2 as cv
import os
# from PIL import Image, ImageDraw, ImageFont
def parse_xml(filepath, filename):
    """解析xml"""
    tree = ET.parse(os.path.join(filepath, filename))  # 解析读取xml函数
    objects = []
    imgname = []
    coordinate = []
    for xml_name in tree.findall('filename'):
        # print('xml_name: ', xml_name)
        # print('xml_name.text: ', xml_name.text)
        # img_path = os.path.join(pic_path, xml_name.text)
        # xml_name.text = '6771.jpg'
        if xml_name.text.split(".")[-1] == "png":
            xml_name.text = xml_name.text.replace('png','jpg')
        # print('xml_name.text: ', xml_name.text)
        imgname.append(xml_name.text)
    for obj in tree.findall('object'):
        obj_struct = {'name': obj.find('name').text, 'pose': obj.find('pose').text,
                      'truncated': int(obj.find('truncated').text), 'difficult': int(obj.find('difficult').text)}
        bbox = obj.find('bndbox')
        obj_struct['bbox'] = [int(bbox.find('xmin').text),
                              int(bbox.find('ymin').text),
                              int(bbox.find('xmax').text),
                              int(bbox.find('ymax').text)]
        objects.append(obj_struct)
    for obj_one in objects:
        xmin = int(obj_one['bbox'][0])
        ymin = int(obj_one['bbox'][1])
        xmax = int(obj_one['bbox'][2])
        ymax = int(obj_one['bbox'][3])
        label = obj_one['name']
        if label == 'face':
            continue
        coordinate.append([xmin,ymin,xmax,ymax,label])
    if len(coordinate) < 1: # remove bad data
        print("imgname: ", imgname)
    return coordinate, imgname

def get_bbox(size, box):
    # Convert xyxy box to YOLOv5 xywh box
    dw = 1. / size[0]
    dh = 1. / size[1]
    xc = (box[0] + box[2])*0.5*dw
    yc = (box[1] + box[3])*0.5*dh
    w = (box[2]-box[0])*dw
    h = (box[3]-box[1])*dh
    return xc, yc, w, h

def xml2coco(path):
    xmldir   = os.path.join(path, 'badbehavior', 'Annotations')
    imgdir   = os.path.join(path, 'badbehavior', 'images')
    labeldir = os.path.join(path, 'badbehavior', 'label')
    # classid: 0-phone, 1-smoke, 2-drink.
    for xmlname in os.listdir(xmldir):
        labelname = os.path.join(labeldir, xmlname.replace('xml', 'txt'))
        labelfile = open(labelname, 'w+')
        tree = ET.parse(os.path.join(xmldir, xmlname))  # 解析读取xml函数
        size = tree.find('size')
        w = int(size.find('width').text)
        h = int(size.find('height').text)
        # print('w: ', w)
        # print('h: ', h)
        for obj in tree.findall('object'):
            obj_struct = {'name': obj.find('name').text, 'pose': obj.find('pose').text,
                          'truncated': int(obj.find('truncated').text), 'difficult': int(obj.find('difficult').text)}
            bbox = obj.find('bndbox')
            obj_struct['bbox'] = [int(bbox.find('xmin').text),
                                  int(bbox.find('ymin').text),
                                  int(bbox.find('xmax').text),
                                  int(bbox.find('ymax').text)]
            xywh = get_bbox((w,h), obj_struct['bbox'])
            # print('xywh: ', xywh)
            if obj_struct['name'] == 'phone':
                classid = 0
            elif obj_struct['name'] == 'smoke':
                classid = 1
            elif obj_struct['name'] == 'drink':
                classid = 2
            else:
                continue
            info = f"{classid} {' '.join(f'{x:.6f}' for x in xywh)}\n"
            # print('info: ', info)
            labelfile.write(info)
        labelfile.close()

def vis_xml(objects, imgdir, imgname):
    """可视化"""
    plotdir = os.path.join(path, 'driver_behavior', 'VOCData', 'plot')
    for _, filename in enumerate(imgname):
        imgpath = os.path.join(imgdir, filename)
        img = cv.imread(imgpath)
        for obj in objects:
            xmin = obj[0]
            ymin = obj[1]
            xmax = obj[2]
            ymax = obj[3]
            label = obj[4]
            cv.rectangle(img, (xmin,ymin), (xmax, ymax), (0, 255, 213), 3)
            cv.putText(img, label, (xmin, ymin), cv.FONT_HERSHEY_SIMPLEX, 2, (225, 18, 255), 1, cv.LINE_AA)
            cv.imshow('vis', img)
            if cv.waitKey(1) == 27:
                print('27')
        aa = os.path.join(plotdir, filename)
        print('aa: ', aa)
        cv.imwrite(aa, img)

def behavior_show(path):
    xmldir = os.path.join(path, 'driver_behavior', 'VOCData', 'Annotations')
    imgdir = os.path.join(path, 'driver_behavior', 'VOCData', 'images')
    k = 0
    for xmlname in os.listdir(xmldir):
        # xmlname = '6771.xml'
        # print('xmlname: ', xmlname)
        # k = k + 1
        # if (k>1):
        #     break
        # parse xml
        objinfo, imgname = parse_xml(xmldir, xmlname)
        # vis xml
        vis_xml(objinfo, imgdir, imgname) 

def phone_csv_show(path):
    csvdir = os.path.join(path, 'phone')
    imgdir = os.path.join(path, 'phone', 'positive')
    plotdir = os.path.join(path, 'phone', 'plot')
    # os.mkdir(plotdir)
    csvfile = open(os.path.join(csvdir, 'labels.csv'), 'r')
    lines = csv.reader(csvfile)
    # print('len: ', type(lines))
    # print('line0: ', lines[0])
    next(lines) # [filename, width, height, xmin, ymin, xmax, ymax, class] 
    # next(lines) # [filename, width, height, xmin, ymin, xmax, ymax, class] 
    annos = []
    for line in lines:
        annos.append(line)
    line = annos[0]
    prename = line[0]
    print('prename: ', prename)
    imgpath = os.path.join(imgdir, prename)
    img = cv.imread(imgpath)
    xmin    = int(line[3])
    ymin    = int(line[4])
    xmax    = int(line[5])
    ymax    = int(line[6])
    # print('xmin: ', type(xmin))
    # print('ymin: ', type(ymin))
    # print('xmax: ', type(xmax))
    # print('ymax: ', type(ymax))
    cv.rectangle(img, (xmin,ymin), (xmax, ymax), (0, 255, 213), 3)
    for k in range(1, len(annos)):
        line = annos[k]
        curname = line[0]
        print('curname: ', curname)
        xmin    = int(line[3])
        ymin    = int(line[4])
        xmax    = int(line[5])
        ymax    = int(line[6])
        if curname == prename:
            cv.rectangle(img, (xmin,ymin), (xmax, ymax), (0, 255, 213), 3)
        else:
            aa = os.path.join(plotdir, prename)
            print('aa: ', aa)
            cv.imwrite(aa, img)
            imgpath = os.path.join(imgdir, curname)
            img = cv.imread(imgpath)
            cv.rectangle(img, (xmin,ymin), (xmax, ymax), (0, 255, 213), 3)
            prename = curname
        cv.imshow('vis', img)
        if cv.waitKey(1) == 27:
            print('27')
    csvfile.close()

def csv2coco(path):
    csvdir = os.path.join(path, 'badbehavior')
    imgdir = os.path.join(path, 'badbehavior', 'positive')
    csvfile = open(os.path.join(csvdir, 'labels.csv'), 'r')
    lines = csv.reader(csvfile)
    next(lines) # [filename, width, height, xmin, ymin, xmax, ymax, class] 
    labeldir = os.path.join(path, 'badbehavior', 'label')
    # classid: 0-phone, 1-smoke, 2-drink.
    imgs = []
    for imgname in os.listdir(imgdir):
        imgs.append(imgname)
    print('imgs len: ', len(imgs))
    for line in lines:
        imgname = line[0]
        print('imgname: ', imgname)
        if line[0] not in imgs:
            print(f"{line[0]} is not in imgs.")
        else:
            # labelname = os.path.join(labeldir, imgname.replace('xml', 'txt'))
            labelname = os.path.join(labeldir, imgname.split('.')[0]+'.txt')
            labelfile = open(labelname, 'a+')
            w       = int(line[1])
            h       = int(line[2])
            xmin    = int(line[3])
            ymin    = int(line[4])
            xmax    = int(line[5])
            ymax    = int(line[6])
            bbox = [xmin, ymin, xmax, ymax]
            classid = 0 # phone
            xywh = get_bbox((w,h), bbox)
            info = f"{classid} {' '.join(f'{x:.6f}' for x in xywh)}\n"
            # print('info: ', info)
            labelfile.write(info)
            labelfile.close()

if __name__ == "__main__":
    path = os.path.dirname(os.path.realpath(__file__))
    # behavior_xml_show(path)
    # phone_csv_show(path)
    # xml2coco(path)
    csv2coco(path)

注意，删除标注数据少的，或者错误的；人脸目标不计算在内；然后将筛选好的数据集整合到一起；

 

 

参考

1. aistudio.baidu.com；

2. XML文件解析--Python_python 解析xml_牧子川的博客-CSDN博客；

完