目标检测xml文件实现mixup数据增强（修改文件路径直接能用，非常方便）

RooKiChen

已于 2022-07-17 22:31:20 修改

阅读量1k

点赞数 1

分类专栏：深度学习文章标签：目标检测 python 人工智能

于 2022-07-14 12:09:07 首次发布

本文为博主原创文章，未经博主允许不得转载

本文链接：https://blog.csdn.net/RooKichenn/article/details/125781646

版权

深度学习专栏收录该内容

23 篇文章 1 订阅

订阅专栏

在这里，先给出原博主的代码：https://blog.csdn.net/qq_36756866/article/details/106299705
这个代码有一些bug，并且生成数据的数量并不能任意改变，于是我自己修改了一部分代码，实现生成任意数据量的功能。使用只需修改文件路径即可，非常方便，不会有任何bug。代码如下：

# coding:utf-8

import cv2
import os
import random
import numpy as np
import xml.etree.ElementTree as ET
import xml.dom.minidom

img_path = '/root/wu_datasets/ReCurrentPapper/data/DOTA/DOTAship/VOCdevkit/VOC2012/JPEGImages/'           # 原始图片文件夹路径
save_path = '/root/wu_datasets/ReCurrentPapper/data/DOTA_mixup/VOCdevkit/VOC2012/JPEGImages/'       # mixup的图片文件夹路径
xml_path = '/root/wu_datasets/ReCurrentPapper/data/DOTA/DOTAship/VOCdevkit/VOC2012/Annotations/'           # 原始图片对应的标注文件xml文件夹的路径
save_xml = '/root/wu_datasets/ReCurrentPapper/data/DOTA_mixup/VOCdevkit/VOC2012//Annotations/'        # mixup的图片对应的标注文件xml的文件夹路径
img_names = os.listdir(img_path)
img_num = len(img_names)
#print(img_names)
# sum 代表生成数据的总和
sum = []
# 这里表示比原先数据集多10倍
for i in range(10):
    for i in range(0, img_num):
        sum.append(i)
# print(sum)
cnt = 0
for k in sum:
    cnt += 1
    print(cnt)
    imgname = img_names[k]
    imgpath = img_path + imgname
    img = cv2.imread(imgpath)
    img_h, img_w = img.shape[0], img.shape[1]
    # print(img_h,img_w)

    i = random.randint(0, img_num - 1)
    add_path = img_path + img_names[i]
    addimg = cv2.imread(add_path)
    add_h, add_w = addimg.shape[0], addimg.shape[1]
    if add_h != img_h or add_w != img_w:
        print('resize!')
        addimg = cv2.resize(addimg, (img_w, img_h), interpolation=cv2.INTER_LINEAR)
    scale_h, scale_w = img_h / add_h, img_w / add_w

    lam = np.random.beta(1.5, 1.5)
    # print(lam)
    mixed_img = lam * img + (1 - lam) * addimg
    save_img = save_path + str(cnt) + '.jpg'
    cv2.imwrite(save_img, mixed_img)
    # print(save_img)

    # print(imgname, img_names[i])
    if imgname != img_names[i]:
        xmlfile1 = xml_path + imgname[:-4] + '.xml'
        xmlfile2 = xml_path + img_names[i][:-4] + '.xml'
        # print(xmlfile1,xmlfile2)

        tree1 = ET.parse(xmlfile1)
        tree2 = ET.parse(xmlfile2)

        doc = xml.dom.minidom.Document()
        root = doc.createElement("annotation")
        doc.appendChild(root)


        for folds in tree1.findall("folder"):
            folder = doc.createElement("folder")
            folder.appendChild(doc.createTextNode(str(folds.text)))
            root.appendChild(folder)
        for filenames in tree1.findall("filename"):
            filename = doc.createElement("filename")
            filename.appendChild(doc.createTextNode("{}.jpg".format(cnt)))
            root.appendChild(filename)
        for paths in tree1.findall("path"):
            path = doc.createElement("path")
            path.appendChild(doc.createTextNode(str(paths.text)))
            root.appendChild(path)
        for sources in tree1.findall("source"):
            source = doc.createElement("source")
            database = doc.createElement("database")
            database.appendChild(doc.createTextNode(str("Unknow")))
            source.appendChild(database)
            root.appendChild(source)
        for sizes in tree1.findall("size"):
            size = doc.createElement("size")
            width = doc.createElement("width")
            height = doc.createElement("height")
            depth = doc.createElement("depth")
            width.appendChild(doc.createTextNode(str(img_w)))
            height.appendChild(doc.createTextNode(str(img_h)))
            depth.appendChild(doc.createTextNode(str(3)))
            size.appendChild(width)
            size.appendChild(height)
            size.appendChild(depth)
            root.appendChild(size)

        nodeframe = doc.createElement("frame")
        nodeframe.appendChild(doc.createTextNode(imgname[:-4]))

        objects = []

        for obj in tree1.findall("object"):
            obj_struct = {}
            obj_struct["name"] = obj.find("name").text
            obj_struct["pose"] = obj.find("pose").text
            obj_struct["truncated"] = obj.find("truncated").text
            obj_struct["difficult"] = 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 in tree2.findall("object"):
            obj_struct = {}
            obj_struct["name"] = obj.find("name").text
            obj_struct["pose"] = obj.find("pose").text
            obj_struct["truncated"] = obj.find("truncated").text
            obj_struct["difficult"] = obj.find("difficult").text          # 有的版本的labelImg改参数为小写difficult
            bbox = obj.find("bndbox")
            obj_struct["bbox"] = [int(int(bbox.find("xmin").text) * scale_w),
                                  int(int(bbox.find("ymin").text) * scale_h),
                                  int(int(bbox.find("xmax").text) * scale_w),
                                  int(int(bbox.find("ymax").text) * scale_h)]
            objects.append(obj_struct)

        for obj in objects:
            nodeobject = doc.createElement("object")
            nodename = doc.createElement("name")
            nodepose = doc.createElement("pose")
            nodetruncated = doc.createElement("truncated")
            nodedifficult = doc.createElement("difficult")
            nodebndbox = doc.createElement("bndbox")
            nodexmin = doc.createElement("xmin")
            nodeymin = doc.createElement("ymin")
            nodexmax = doc.createElement("xmax")
            nodeymax = doc.createElement("ymax")
            nodename.appendChild(doc.createTextNode(obj["name"]))
            nodepose.appendChild(doc.createTextNode(obj["pose"]))
            nodetruncated.appendChild(doc.createTextNode(obj["truncated"]))
            nodedifficult.appendChild(doc.createTextNode(obj["difficult"]))
            nodexmin.appendChild(doc.createTextNode(str(obj["bbox"][0])))
            nodeymin.appendChild(doc.createTextNode(str(obj["bbox"][1])))
            nodexmax.appendChild(doc.createTextNode(str(obj["bbox"][2])))
            nodeymax.appendChild(doc.createTextNode(str(obj["bbox"][3])))

            nodebndbox.appendChild(nodexmin)
            nodebndbox.appendChild(nodeymin)
            nodebndbox.appendChild(nodexmax)
            nodebndbox.appendChild(nodeymax)

            nodeobject.appendChild(nodename)
            nodeobject.appendChild(nodepose)
            nodeobject.appendChild(nodetruncated)
            nodeobject.appendChild(nodedifficult)
            nodeobject.appendChild(nodebndbox)

            root.appendChild(nodeobject)

        fp = open(save_xml + str(cnt) + ".xml", "w")
        doc.writexml(fp, indent='\t', addindent='\t', newl='\n', encoding="utf-8")
        fp.close()

    else:
        xmlfile1 = xml_path + imgname[:-4] + '.xml'
        # print(xmlfile1)
        tree1 = ET.parse(xmlfile1)

        doc = xml.dom.minidom.Document()
        root = doc.createElement("annotation")


        doc.appendChild(root)

        for folds in tree1.findall("folder"):
            folder=doc.createElement("folder")
            folder.appendChild(doc.createTextNode(str(folds.text)))
            root.appendChild(folder)
        for filenames in tree1.findall("filename"):
            filename=doc.createElement("filename")
            filename.appendChild(doc.createTextNode("{}.jpg".format(cnt)))
            root.appendChild(filename)
        for paths in tree1.findall("path"):
            path = doc.createElement("path")
            path.appendChild(doc.createTextNode(str(paths.text)))
            root.appendChild(path)
        for sources in tree1.findall("source"):
            source = doc.createElement("source")
            database = doc.createElement("database")
            database.appendChild(doc.createTextNode(str("Unknow")))
            source.appendChild(database)
            root.appendChild(source)
        for sizes in tree1.findall("size"):
            size = doc.createElement("size")
            width = doc.createElement("width")
            height = doc.createElement("height")
            depth = doc.createElement("depth")
            width.appendChild(doc.createTextNode(str(img_w)))
            height.appendChild(doc.createTextNode(str(img_h)))
            depth.appendChild(doc.createTextNode(str(3)))
            size.appendChild(width)
            size.appendChild(height)
            size.appendChild(depth)
            root.appendChild(size)


        nodeframe = doc.createElement("frame")
        nodeframe.appendChild(doc.createTextNode(imgname[:-4]))
        objects = []

        for obj in tree1.findall("object"):
            obj_struct = {}
            obj_struct["name"] = obj.find("name").text
            obj_struct["pose"] = obj.find("pose").text
            obj_struct["truncated"] = obj.find("truncated").text
            obj_struct["difficult"] = 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 in objects:
            nodeobject = doc.createElement("object")
            nodename = doc.createElement("name")
            nodepose = doc.createElement("pose")
            nodetruncated = doc.createElement("truncated")
            nodedifficult = doc.createElement("difficult")
            nodebndbox = doc.createElement("bndbox")
            nodexmin = doc.createElement("xmin")
            nodeymin = doc.createElement("ymin")
            nodexmax = doc.createElement("xmax")
            nodeymax = doc.createElement("ymax")
            nodename.appendChild(doc.createTextNode(obj["name"]))
            nodepose.appendChild(doc.createTextNode(obj["pose"]))
            nodetruncated.appendChild(doc.createTextNode(obj["truncated"]))
            nodedifficult.appendChild(doc.createTextNode(obj["difficult"]))
            nodexmin.appendChild(doc.createTextNode(str(obj["bbox"][0])))
            nodeymin.appendChild(doc.createTextNode(str(obj["bbox"][1])))
            nodexmax.appendChild(doc.createTextNode(str(obj["bbox"][2])))
            nodeymax.appendChild(doc.createTextNode(str(obj["bbox"][3])))

            nodebndbox.appendChild(nodexmin)
            nodebndbox.appendChild(nodeymin)
            nodebndbox.appendChild(nodexmax)
            nodebndbox.appendChild(nodeymax)

            nodeobject.appendChild(nodename)
            nodeobject.appendChild(nodepose)
            nodeobject.appendChild(nodetruncated)
            nodeobject.appendChild(nodedifficult)
            nodeobject.appendChild(nodebndbox)

            root.appendChild(nodeobject)

        fp = open(save_xml + str(cnt) + ".xml", "w")
        doc.writexml(fp, indent='\t', addindent='\t', newl='\n', encoding="utf-8")
        fp.close()