Python uses the xml format label file under the folder to crop out the targets in the picture in batches (the number of folders, picture names, target boxes is unlimited, and line-by-line comments)

During this period of time, a small task was to extract the hand objects in the picture, and then I used labelimg marking software to mark some pictures (generating xml label files), and wrote a small script to remove these targets from the original picture. Cut out in the middle, the effect is shown below, and it feels okay.

Original picture (example 1):

cropping effect (example 1):

original picture (example 2):

cropping effect (example 2, here is an empty hand for marking, so it is not cropped):

Complete code

This code solves the following problems:

1. The same xml file has multiple target boxes;
2. The picture and the xml file are in different or the same folder;
3. Some pictures are not marked, and the corresponding xml file does not exist;
4. There are non-pictures in the picture folder ( png, jpg) files;
5. Irregular naming of pictures

import cv2
import xml.etree.ElementTree as ET
import os 

img_path = r'D:\Dataset\HoldingObject\11\color_S000C000P001R000A011' #图片路径
xml_path = r'D:\Dataset\HoldingObject\11\color_S000C000P001R000A011' #标签路径
obj_img_path = r'D:\Dataset\HoldingObject\11\11-1'   #目标裁剪图片存放路径

for img_file in os.listdir(img_path):    #遍历图片文件夹
    if img_file[-4:] in ['.png', '.jpg']:    #判断文件是否为图片格式
        img_filename = os.path.join(img_path, img_file)  #将图片路径与图片名进行拼接
        img_cv = cv2.imread(img_filename)  #读取图片
        
        img_name = (os.path.splitext(img_file)[0])  #分割出图片名，如“000.png” 图片名为“000”
        xml_name = xml_path + '\\' + '%s.xml'%img_name  #利用标签路径、图片名、xml后缀拼接出完整的标签路径名
        
        if os.path.exists(xml_name):  #判断与图片同名的标签是否存在，因为图片不一定每张都打标
            root = ET.parse(xml_name).getroot() #利用ET读取xml文件
            count = 0 #目标框个数统计，防止目标文件覆盖
            for obj in root.iter('object'):  #遍历所有目标框
                name = obj.find('name').text   #获取目标框名称，即label名
            
                xmlbox = obj.find('bndbox')   #找到框目标
                x0 = xmlbox.find('xmin').text  #将框目标的四个顶点坐标取出
                y0 = xmlbox.find('ymin').text
                x1 = xmlbox.find('xmax').text
                y1 = xmlbox.find('ymax').text
                
                obj_img = img_cv[int(y0):int(y1), int(x0):int(x1)]  #cv2裁剪出目标框中的图片
                
                cv2.imwrite(obj_img_path + '\\' + '%s_%s'%(img_name, count) + '.jpg', obj_img)  #保存裁剪图片
                count += 1 #目标框统计值自增1