前言：对于航拍、遥感影像数据集而言，此类数据集包含较多目标，且目标相对较小，直接进行目标检测往往效果不佳，此时对大分辨率图像做切割就是一个不错选择，也扩充了数据量。

~~~本文的切割思路是先将xml文件转换为dota格式的txt文件，然后利用dota切割方法对图像进行切割，最后再将切割后的文件转换为yolo格式的txt文件~~~

1、xml文件转换为dota格式的txt文件

#  VOC标注格式为左上角与右下角坐标
#  这个文件直接将标注框左上角与右下角坐标转换为四个角的坐标（顺时针）
import xml.etree.ElementTree as ET
import pickle
import os
import os.path
from os import listdir, getcwd
from os.path import join
sets = ['train', 'test', 'val']
classes = ['num0', 'num1', 'num2', 'num3', 'num4', 'num5', 'num6', 'num7', 'num8', 'num9']def convert(size, box):dw = 1. / size[0]dh = 1. / size[1]x = (box[0] + box[1]) / 2.0y = (box[2] + box[3]) / 2.0w = box[1] - box[0]h = box[3] - box[2]x = x * dww = w * dwy = y * dhh = h * dhreturn (x, y, w, h)
def convert_annotation(image_id):if os.path.isfile('datasets/mydata/Annotations/%s.xml' % (image_id)):in_file = open('datasets/mydata/Annotations/%s.xml' % (image_id))out_file = open('datasets/mydata/labels/%s.txt' % (image_id), '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').textcls = obj.find('name').textif cls not in classes or int(difficult) == 1:continue# cls_id = classes.index(cls)xmlbox = obj.find('bndbox')# 注意dota数据格式四个角点的坐标是顺时针分布的x0 = int(float(xmlbox.find('xmin').text))y0 = int(float(xmlbox.find('ymin').text))x1 = int(float(xmlbox.find('xmax').text))y1 = int(float(xmlbox.find('ymin').text))x2 = int(float(xmlbox.find('xmax').text))y2 = int(float(xmlbox.find('ymax').text))x3 = int(float(xmlbox.find('xmin').text))y3 = int(float(xmlbox.find('ymax').text))out_file.write("{} {} {} {} {} {} {} {} {} {}\n".format(x0, y0, x1, y1, x2, y2, x3, y3, cls, difficult))
wd = getcwd()
print(wd)
for image_set in sets:if not os.path.exists('datasets/mydata/labels/'):os.makedirs('datasets/mydata/labels/')image_ids = open('datasets/mydata/%s.txt' % (image_set)).read().strip().split()list_file = open('datasets/mydata/%s.txt' % (image_set), 'w')for image_id in image_ids:list_file.write('datasets/mydata/images/%s.jpg\n' % (image_id))convert_annotation(image_id)list_file.close()#数据存放格式
#|——根目录
#    |——datasets
#        |——dota_images
#        |——dota_labels
#        |——yolo_labels
#        |——mydata
#            |——Annotations
#            |——images
#            |——labels
#               test.txt 
#               train.txt 
#               val.txt 

2、图像切割

import cv2
import os#  图像宽不足裁剪宽度,填充至裁剪宽度
def fill_right(img, size_w):size = img.shape#  填充值为数据集均值img_fill_right = cv2.copyMakeBorder(img, 0, 0, 0, size_w - size[1],cv2.BORDER_CONSTANT, value = (107, 113, 115))return img_fill_right#  图像高不足裁剪高度,填充至裁剪高度
def fill_bottom(img, size_h):size = img.shapeimg_fill_bottom = cv2.copyMakeBorder(img, 0, size_h - size[0], 0, 0,cv2.BORDER_CONSTANT, value = (107, 113, 115))return img_fill_bottom#  图像宽高不足裁剪宽高度,填充至裁剪宽高度
def fill_right_bottom(img, size_w, size_h):size = img.shapeimg_fill_right_bottom = cv2.copyMakeBorder(img, 0, size_h - size[0], 0, size_w - size[1],cv2.BORDER_CONSTANT, value = (107, 113, 115))return img_fill_right_bottom#  图像切割
#  img_floder 图像文件夹
#  out_img_floder 图像切割输出文件夹
#  size_w 切割图像宽
#  size_h 切割图像高
#  step 切割步长
#  不管输入图片是什么格式，切割出来的图片都为.bmp格式（可自行设置）
def image_split(img_floder, out_img_floder, size_w, size_h, step):img_list = os.listdir(img_floder)# print(img_list)count = 0for img_name in img_list:number = 0#  去除.png后缀name = img_name[:-4]img = cv2.imread(img_floder + "\\" + img_name)size = img.shape#  若图像宽高大于切割宽高if size[0] >= size_h and size[1] >= size_w:count = count + 1for h in range(0, size[0] - 1, step):start_h = hfor w in range(0, size[1] - 1, step):start_w = wend_h = start_h + size_hif end_h > size[0]:start_h = size[0] - size_hend_h = start_h + size_hend_w = start_w + size_wif end_w > size[1]:start_w = size[1] - size_wend_w = start_w + size_wcropped = img[start_h : end_h, start_w : end_w]#  用起始坐标来命名切割得到的图像，为的是方便后续标签数据抓取name_img = name + '_'+ str(start_h) +'_' + str(start_w)cv2.imwrite('{}/{}.bmp'.format(out_img_floder, name_img), cropped)number = number + 1#  若图像高大于切割高,但宽小于切割宽elif size[0] >= size_h and size[1] < size_w:print('图片{}需要在右面补齐'.format(name))count = count + 1img0 = fill_right(img, size_w)for h in range(0, size[0] - 1, step):start_h = hstart_w = 0end_h = start_h + size_hif end_h > size[0]:start_h = size[0] - size_hend_h = start_h + size_hend_w = start_w + size_wcropped = img0[start_h : end_h, start_w : end_w]name_img = name + '_' + str(start_h) + '_' + str(start_w)cv2.imwrite('{}/{}.bmp'.format(out_img_floder, name_img), cropped)number = number + 1#  若图像宽大于切割宽,但高小于切割高elif size[0] < size_h and size[1] >= size_w:count = count + 1print('图片{}需要在下面补齐'.format(name))img0 = fill_bottom(img, size_h)for w in range(0, size[1] - 1, step):start_h = 0start_w = wend_w = start_w + size_wif end_w > size[1]:start_w = size[1] - size_wend_w = start_w + size_wend_h = start_h + size_hcropped = img0[start_h : end_h, start_w : end_w]name_img = name + '_'+ str(start_h) +'_' + str(start_w)cv2.imwrite('{}/{}.bmp'.format(out_img_floder, name_img), cropped)number = number + 1#  若图像宽高小于切割宽高elif size[0] < size_h and size[1] < size_w:count = count + 1print('图片{}需要在下面和右面补齐'.format(name))img0 = fill_right_bottom(img,  size_w, size_h)cropped = img0[0 : size_h, 0 : size_w]name_img = name + '_'+ '0' +'_' + '0'cv2.imwrite('{}/{}.bmp'.format(out_img_floder, name_img), cropped)number = number + 1print('{}.bmp切割成{}张.'.format(name,number))print('共完成{}张图片'.format(count))#  txt切割
#  out_img_floder 图像切割输出文件夹
#  txt_floder txt文件夹
#  out_txt_floder txt切割输出文件夹
#  size_w 切割图像宽
#  size_h 切割图像高
def txt_split(out_img_floder, txt_floder, out_txt_floder, size_h, size_w):img_list = os.listdir(out_img_floder)for img_name in img_list:#  去除.png后缀name = img_name[:-4]#  得到原图像(也即txt)索引 + 切割高 + 切割宽name_list = name.split('_')# 易报错之处，此处需要跟据图片的具体名称来改txt_name = name_list[0] + '_' + name_list[1]h = int(name_list[2])w = int(name_list[3])# txt_name = name_list[0]# h = int(name_list[1])# w = int(name_list[2])txtpath = txt_floder + "\\" + txt_name + '.txt'out_txt_path = out_txt_floder + "\\" + name + '.txt'f = open(out_txt_path, 'a')#  打开txt文件with open(txtpath, 'r') as f_in:lines = f_in.readlines()#  逐行读取for line in lines:splitline = line.split(' ')label = splitline[8]difficult = splitline[9]x1 = int(float(splitline[0]))y1 = int(float(splitline[1]))x2 = int(float(splitline[2]))y2 = int(float(splitline[3]))x3 = int(float(splitline[4]))y3 = int(float(splitline[5]))x4 = int(float(splitline[6]))y4 = int(float(splitline[7]))if w <= x1 <= w + size_w and w <= x2 <= w + size_w and \w <= x3 <= w + size_w and w <= x4 <= w + size_w and \h <= y1 <= h + size_h and h <= y2 <= h + size_h and \h <= y3 <= h + size_h and h <= y4 <= h + size_h:f.write('{} {} {} {} {} {} {} {} {} {}'.format(int(x1 - w),int(y1 - h), int(x2 - w), int(y2 - h), int(x3 - w),int(y3 - h), int(x4 - w), int(y4 - h),label, difficult))f.close()# print('{}.txt切割完成.'.format(name))print('txt切割完成')#  图像数据集文件夹
img_floder = r'D:\cnn\yolov5-6.10\datasets\mydata\images'
#  切割得到的图像数据集存放文件夹
out_img_floder = r'D:\cnn\yolov5-6.10\datasets\dota_images'
#  txt数据集文件夹
txt_floder = r'D:\cnn\yolov5-6.10\datasets\mydata\labels'
#  切割后数据集的标签文件存放文件夹
out_txt_floder = r'D:\cnn\yolov5-6.10\datasets\dota_labels'
#  此为切割出来的图像宽
size_w = 640
#  此为切割出来的图像高
size_h = 640
#  切割步长,重叠度为size_w - step
step = 540image_split(img_floder, out_img_floder, size_w, size_h, step)
txt_split(out_img_floder, txt_floder, out_txt_floder, size_h, size_w)# 清除未包含标注框的图片和txt文件
txt_lists = os.listdir(out_txt_floder) # 读取输出的txt文件
j = 0
for txt_list in txt_lists:# print(txt_list)name = txt_list[:-4]  # 去除.txt后缀# print(name)data = open(r'D:\cnn\yolov5-6.10\datasets\dota_labels\{}'.format(txt_list)).read()if len(data) == 0:j = j + 1# print("{}文件为空！".format(txt_list))# os.remove('D:\cnn\yolov5-6.10\datasets\dota_labels')path1 = r'D:\cnn\yolov5-6.10\datasets\dota_labels\{}'.format(txt_list)os.remove(path1) # 删除空的txt文件path2 = r'D:\cnn\yolov5-6.10\datasets\dota_images\{}'.format(name + '.bmp')os.remove(path2) # 删除对应的没有标注框的图片
print('共{}个txt文件为空'.format(j))#数据存放格式
#|——根目录
#    |——datasets
#        |——dota_images
#        |——dota_labels
#        |——yolo_labels
#        |——mydata
#            |——Annotations
#            |——images
#            |——labels
#               test.txt 
#               train.txt 
#               val.txt 

运行完后会在dota_images和dota_labels文件夹中生成切割后的图片和txt文件。

3、将dota格式的txt文件转换为yolo格式的txt文件

# 在网上找了很久没找到转化的代码，索性就自己写了一个import osclasses = ['num0', 'num1', 'num2', 'num3', 'num4', 'num5', 'num6', 'num7', 'num8', 'num9']
path = "D:\cnn\yolov5-6.10\datasets\dota_labels"  # dota格式txt文件所在目录
yolo_label = 'D:\cnn\yolov5-6.10\datasets\yolo_labels' # yolo格式txt文件保存目录label_lists = os.listdir(path)  # 得到文件夹下的所有文件名称
for label_list in label_lists:  # 遍历文件夹out_txt_path = yolo_label + "\\" + label_list # 构造输出绝对路径，label_list为txt文件名称，即保证文件命名不变position = path + '\\' + label_list  # 构造绝对路径，"\\"，其中一个'\'为转义符# print(position)with open(position, 'r') as f:  # 打开文件lines = f.readlines()  # 逐行读取文件file = open(out_txt_path, 'w') # 写入nums = len(lines) # 总共有多少行for i in range(nums): # 遍历每一行values = lines[i]value = values.split(' ')# 因为dota格式四个角为顺时针，转为yolo格式需要取出其左上角和右下角xmin = float(value[0])ymin = float(value[1])xmax = float(value[4])ymax = float(value[5])cls = value[8]cls_id = classes.index(cls)d = 1. / 640 # 归一化处理，此处切割出来的图片宽高都是640file.write('{} {} {} {} {}\n'.format(cls_id, float(xmin * d), float(ymin * d), float(xmax * d), float(ymax * d)))

tips：一定要记得归一化处理，因为yolo格式是要求归一化后的坐标！！！ 

至此，就得到了用于yolo训练的切割数据集！！！ 

跑程序的时候发现这种yolo格式不对（目前还不知道什么原因），继续往下看~~~

4、将yolo格式的txt文件转换为xml文件，此时可以用labelimg查看切割效果

import os
import glob
from PIL import Imageyolo_labels = r'D:\cnn\yolov5-6.10\datasets\yolo_labels'  # yolo格式下的存放txt标注文件的文件夹
# 这里将图片和xml文件存放在一个文件夹中，方便后续使用labelimg查看
xml_labels = r'D:\cnn\yolov5-6.10\datasets\dota_images'  # 转换后储存xml的文件夹地址
dota_images = r'D:\cnn\yolov5-6.10\datasets\dota_images'  # 存放图片的文件夹labels = ['num0', 'num1', 'num2', 'num3', 'num4', 'num5', 'num6', 'num7', 'num8', 'num9']
src_img_dir = dota_images
src_txt_dir = yolo_labels
src_xml_dir = xml_labels  # 转换后储存xml的文件夹地址img_lists = glob.glob(src_img_dir + '/*.bmp')
img_basenames = []
for item in img_lists:img_basenames.append(os.path.basename(item))  # os.path.basename返回path最后的文件名img_names = []
for item in img_basenames:temp1, temp2 = os.path.splitext(item)  # os.path.splitext(“文件路径”)  分离文件名与扩展名img_names.append(temp1)total_num = len(img_names)  # 统计当前总共要转换的图片标注数量
count = 0  # 技术变量
for img in img_names:  # 这里的img是不加后缀的图片名称count += 1if count % 1000 == 0:print("当前转换进度{}/{}".format(count, total_num))im = Image.open((src_img_dir + '/' + img + '.bmp'))width, height = im.size# 打开yolo格式下的txt文件gt = open(src_txt_dir + '/' + img + '.txt').read().splitlines()if gt:# 将主干部分写入xml文件中xml_file = open((src_xml_dir + '/' + img + '.xml'), 'w')xml_file.write('\n')xml_file.write('    VOC2007\n')xml_file.write('    ' + str(img) + '.bmp' + '\n')xml_file.write('    \n')xml_file.write('        ' + str(width) + '\n')xml_file.write('        ' + str(height) + '\n')xml_file.write('        3\n')xml_file.write('    \n')# write the region of image on xml filefor img_each_label in gt:spt = img_each_label.split(' ')  # 这里如果txt里面是以逗号‘，’隔开的，那么就改为spt = img_each_label.split(',')。xml_file.write('    \n')xml_file.write('')else:# 将主干部分写入xml文件中xml_file = open((src_xml_dir + '/' + img + '.xml'), 'w')xml_file.write('\n')xml_file.write('    VOC2007\n')xml_file.write('    ' + str(img) + '.bmp' + '\n')xml_file.write('    \n')xml_file.write('        ' + str(width) + '\n')xml_file.write('        ' + str(height) + '\n')xml_file.write('        3\n')xml_file.write('    \n')xml_file.write('')

使用labelimg打开dota_images文件夹就可以查看切割后的效果啦！！！

5、 再将xml格式文件转换为txt格式文件

import xml.etree.ElementTree as ET
import pickle
import os
import os.path
from os import listdir, getcwd
from os.path import join
sets = ['train', 'test', 'val']
classes = ['num0', 'num1', 'num2', 'num3', 'num4', 'num5', 'num6', 'num7', 'num8', 'num9']def convert(size, box):dw = 1. / size[0]dh = 1. / size[1]x = (box[0] + box[1]) / 2.0y = (box[2] + box[3]) / 2.0w = box[1] - box[0]h = box[3] - box[2]x = x * dww = w * dwy = y * dhh = h * dhreturn (x, y, w, h)
def convert_annotation(image_id):if os.path.isfile('datasets/mydata/Annotations/%s.xml' % (image_id)):in_file = open('datasets/mydata/Annotations/%s.xml' % (image_id))out_file = open('datasets/mydata/labels/%s.txt' % (image_id), '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').textcls = obj.find('name').textif cls not in classes or int(difficult) == 1:continuecls_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))bb = convert((w, h), b)out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')
wd = getcwd()
print(wd)
for image_set in sets:if not os.path.exists('datasets/mydata/labels/'):os.makedirs('datasets/mydata/labels/')image_ids = open('datasets/mydata/%s.txt' % (image_set)).read().strip().split()list_file = open('datasets/mydata/%s.txt' % (image_set), 'w')for image_id in image_ids:list_file.write('datasets/mydata/images/%s.jpg\n' % (image_id))convert_annotation(image_id)list_file.close()

至此，就得到了用于yolo训练的切割数据集！！！