高分辨率图像建筑物提取数据集制作

目录结构

/dataset/
    xxxx.tif  # 原始图像
    image-3000
        0.tif # 切割后的tif
        1.tif
        ........
        xxx_0.json # 生成的json文件也放在该文件夹下
        xxx_1.json
        .......
        xxx_0_json # 调用labelme_json_to_dataset 0.json 生成的0_json文件夹
            img.png # 原始图片的png格式
            info.yaml
            label.png # 标签图片
            label_names.txt 
            label_viz.png
        xxx_1_json
        .......
        0.png  # 将json文件夹中的label.png 提取出来
        1.png 
        .......
        label_0.tif # 将上边的png标签文件转换为tif格式

收集数据，高分辨率图像

1. 无人机数据，航空数据等

图像切割，像素大小该为多少？

1. 本数据集平均像素大小（40000*50000） tif格式，LZW压缩方式压缩

2. 要考虑计算机显卡，目标建筑无尽量不被切割等问题，本利使用3000*3000

   # data:2020-01-04
   # user:dean
   # desc:图像切割脚本
   import tifffile as tiff  # 也可使用pillow或opencv 但若图片过大时可能会出问题
   import os
   width = 1500*2   # 切割图像大小
   height = 1500*2  # 切割图像大小
   home = "/media/dean/Document/AI_dataset/DOM/"
   file_name = "裴庄村51-dom"
   image_dir = os.path.join(home,file_name)
   image = os.path.join(image_dir,file_name+".tif")
   target_dir = os.path.join(image_dir,"image-"+str(width))  # 切割后图片存储位置
   if not os.path.exists(target_dir):
       os.mkdir(target_dir)
   img = tiff.imread(image)  # 导入图片
   print("导入图片完成",img.shape) # 原始图片大小
   pic_width = img.shape[1]
   pic_height = img.shape[0]
   row_num = pic_width//width  # 纵向切割数量
   col_num = pic_height // height  # 横向切割数量
   print("开始进行切割，可切割总数为{}".format(col_num*row_num))
   for j in range(col_num):
       for i in range(row_num):
           num = j * row_num + i
           print("正在进行第{}张切割".format(num + 1))
           row = i * width
           row_end = row + width
           col = j * height
           col_end = col + height
           # print(col,col_end,row,row_end)
           cropped = img[col:col_end,row:row_end]
           name = "{}_{}.tif".format(file_name,num)
           image_path = os.path.join(target_dir,name)
           tiff.imsave(image_path, cropped)

标注工具 labelme

1. 使用label标注每张图片

   pip install labelme  # 安装labelme

每张图片标注后会生成对应name.json文件

labelme_json_to_dataset xxx.json 

# data:2020-01-04
# user:dean
# desc:批量将json文件转为 label
import os
dir = r"I:\人工智能数据\DOM\裴庄村51-dom\image-3000"
files = [os.path.join(dir,file) for file in os.listdir(dir) if file.endswith(".json")]
for file in files:
    cmd = "labelme_json_to_dataset {}".format(file)
    print(cmd)
    os.system(cmd)

将所有的json/label.png 提取到统一文件夹

# data:2020-01-04
# user:dean
# desc:将label文件夹中的laebl提取出来
import tifffile as tiff
from PIL import Image
import os
target_dir = r"/media/dean/Document/AI_dataset/DOM/裴庄村51-dom/image-3000"  # json_label 所在的文件夹
files = [os.path.join(target_dir,file) for file in os.listdir(target_dir)]
for i in files:
    if os.path.isdir(i):
        lables = os.listdir(i)
        for file in lables:
            if file == "label.png":
                image_path = os.path.join(i, "label.png")
                imgae = Image.open(image_path)
                parent_dir_name = os.path.basename(os.path.dirname(image_path))
                new_name = "{}.png".format(parent_dir_name.split("_")[1])
                imgae.save(os.path.join(target_dir,new_name))
                print("第{}个文件夹".format(i))
                break;

将所有的label.png转换为tif格式 并转换为单通道黑白照片

# coding:utf-8
# file: change_format.py
# author: Dean
# contact: 1028968939@qq.com
# time: 2020/1/4 20:41
# desc: 将png 标签转化为单通道 黑白标签 并转化为tif
import os
from PIL import Image
threshold = 0
table = []
for i in range(256):
    if i > threshold:
        table.append(255)
    else:
        table.append(0)
target_dir = r"/media/dean/Document/AI_dataset/DOM/裴庄村51-dom/image-3000"
files = [os.path.join(target_dir,file) for file in os.listdir(target_dir) if file.endswith(".png")]
for file in files:
    image_file_name = os.path.basename(file)
    num = image_file_name.split(".")[0]
    image_file = Image.open(file)  # open colour image
    # image_file = image_file.convert('L') # convert image to black and white
    image_file = image_file.point(table, '1')
    new_file = os.path.join(target_dir,"{}.tif".format(num))
    image_file.save(new_file)
    print(new_file)