labelme语义分割标注工具直接生成图像的改进方法
使用过语义分割模型的人应该对labelme工具不会陌生，labelme工具允许我们以正多边形拟合的方式对图像中的目标区域进行标注，标注的结果保存在json文件中，json文件和被标注的原始图像一一对应。
但是在后续的使用中，很多语义分割的训练数据要求lable结果是图像的形式，也就是需要将json的结果转换成图像进而进行训练。
在网上看了很多方法将json转成图片，但是往往需要现用labelme得到json再进行转换，也就是两步操作；之前我也是一样的操作；但是最近由于要把标注工作教给很多人，两步的转换增添了工作量，于是考虑直接改动labelme的源码，使得标注结果直接就有图片形式和json格式。
操作如下：
- 安装labelme，推荐使用anaconda安装
# python2
conda create --name=labelme python=2.7
source activate labelme
# conda install -c conda-forge pyside2
conda install pyqt
pip install labelme
# if you'd like to use the latest version. run below:
# pip install git+https://github.com/wkentaro/labelme.git
# python3
conda create --name=labelme python=3.6
source activate labelme
# conda install -c conda-forge pyside2
# conda install pyqt
# pip install pyqt5 # pyqt5 can be installed via pip on python3
pip install labelme
# or you can install everything by conda command
# conda install labelme -c conda-forge
- 安装完以后找到anaconda的安装路径，进入env中的labelme，一下是我的路径，参考。
D:\anaconda\envs\labelme\Lib\site-packages\labelme
找到其中的label_file.py文件，修改如下。
import base64
import io
import json
import os.path as osp
import PIL.Image
from labelme._version import __version__
from labelme.logger import logger
from labelme import PY2
from labelme import QT4
from labelme import utils
MODE = 'BW' # 保存的mask图片类型，二值图为BW，彩图为RGB
import numpy as np
from labelme import utils
from labelme.utils.draw import label_colormap
import PIL.Image
import cv2
import os
class LabelFileError(Exception):
pass
class LabelFile(object):
suffix = '.json'
def __init__(self, filename=None):
self.shapes = ()
self.imagePath = None
self.imageData = None
if filename is not None:
self.load(filename)
self.filename = filename
@staticmethod
def load_image_file(filename):
try:
image_pil = PIL.Image.open(filename)
##################################
print("filename is :", filename)
except IOError:
logger.error('Failed opening image file: {}'.format(filename))
return
# apply orientation to image according to exif
image_pil = utils.apply_exif_orientation(image_pil)
with io.BytesIO() as f:
ext = osp.splitext(filename)[1].lower()
if PY2 and QT4:
format = 'PNG'
elif ext in ['.jpg', '.jpeg']:
format = 'JPEG'
else:
format = 'PNG'
image_pil.save(f, format=format)
f.seek(0)
return f.read()
def load(self, filename):
keys = [
'imageData',
'imagePath',
'lineColor',
'fillColor',
'shapes', # polygonal annotations
'flags', # image level flags
'imageHeight',
'imageWidth',
]
try:
with open(filename, 'rb' if PY2 else 'r') as f:
data = json.load(f)
if data['imageData'] is not None:
imageData = base64.b64decode(data['imageData'])
if PY2 and QT4:
imageData = utils.img_data_to_png_data(imageData)
else:
# relative path from label file to relative path from cwd
imagePath = osp.join(osp.dirname(filename), data['imagePath'])
imageData = self.load_image_file(imagePath)
flags = data.get('flags') or {}
imagePath = data['imagePath']
self._check_image_height_and_width(
base64.b64encode(imageData).decode('utf-8'),
data.get('imageHeight'),
data.get('imageWidth'),
)
lineColor = data['lineColor']
fillColor = data['fillColor']
shapes = (
(
s['label'],
s['points'],
s['line_color'],
s['fill_color'],
s.get('shape_type', 'polygon'),
s.get('flags', {}),
)
for s in data['shapes']
)
except Exception as e:
raise LabelFileError(e)
otherData = {}
for key, value in data.items():
if key not in keys:
otherData[key] = value
# Only replace data after everything is loaded.
self.flags = flags
self.shapes = shapes
self.imagePath = imagePath
self.imageData = imageData
self.lineColor = lineColor
self.fillColor = fillColor
self.filename = filename
self.otherData = otherData
@staticmethod
def _check_image_height_and_width(imageData, imageHeight, imageWidth):
img_arr = utils.img_b64_to_arr(imageData)
if imageHeight is not None and img_arr.shape[0] != imageHeight:
logger.error(
'imageHeight does not match with imageData or imagePath, '
'so getting imageHeight from actual image.'
)
imageHeight = img_arr.shape[0]
if imageWidth is not None and img_arr.shape[1] != imageWidth:
logger.error(
'imageWidth does not match with imageData or imagePath, '
'so getting imageWidth from actual image.'
)
imageWidth = img_arr.shape[1]
return imageHeight, imageWidth
'''
保存json和使用json画图并保存，画图原理基本就是创建一个和原图尺寸相同的画布，
然后根据json中描述的正多边形定点的键值对将正多边形内部的区域像素置换成
指定的颜色，保存成新的图片完成标注
'''
def save(
self,
filename,
shapes,
imagePath,
imageHeight,
imageWidth,
imageData=None,
lineColor=None,
fillColor=None,
otherData=None,
flags=None,
):
if imageData is not None:
imageData = base64.b64encode(imageData).decode('utf-8')
imageHeight, imageWidth = self._check_image_height_and_width(
imageData, imageHeight, imageWidth
)
if otherData is None:
otherData = {}
if flags is None:
flags = {}
data = dict(
version=__version__,
flags=flags,
shapes=shapes,
lineColor=lineColor,
fillColor=fillColor,
imagePath=imagePath,
imageData=imageData,
imageHeight=imageHeight,
imageWidth=imageWidth,
)
#################################################
#imageData = base64.b64encode(imageData).decode('utf-8')
img = utils.img_b64_to_arr(imageData)
#print('img array is:',img)
# 将标记的json数据转换为图片
label_name_to_value = {'_background_': 0}
for shape in sorted(data['shapes'], key=lambda x: x['label']):
#print("json shape is:", shape)
label_name = shape['label']
if label_name in label_name_to_value:
label_value = label_name_to_value[label_name]
else:
label_value = len(label_name_to_value)
label_name_to_value[label_name] = label_value
lbl = utils.shapes_to_label(img.shape, data['shapes'], label_name_to_value)
#cv2.imwrite("tower-label.jpg", lbl * 255)
# 保存图片
out_dir = os.path.join(os.path.abspath(os.path.join(filename,os.path.pardir)), 'mask')
#out_dir = r"C:\Users\admin\Desktop\cv22\标注工具\mask-test"
if not osp.exists(out_dir):
os.mkdir(out_dir)
#save_name = os.path.join(out_dir, imagePath.split('.')[0]+'.jpg')
#save_name = "tower.jpg"
save_name = os.path.join(out_dir,imagePath)
print("out_dir is :",out_dir)
print("imagepath is :",imagePath)
print("save_name is :",save_name)
if MODE == 'BW':
lbl_tmp = cv2.resize(lbl, (500,400), interpolation=cv2.INTER_NEAREST) #resize操作
lbl_tmp = 255 - lbl_tmp * 255 #颠倒黑白,将数据分布在0-255范围
lbl_tmp[lbl_tmp>100] = 255
lbl_tmp[lbl_tmp<=100] = 0 #因为我的应用是二分类，所以全图强制二值化
#print("img shape before resize is:",lbl_tmp.shape)
#for i in range(224):
#print(i in lbl_tmp)
#print(lbl_tmp.shape)
#lbl_tmp = lbl_tmp[255-lbl_tmp]
#cv2.imwrite("tower1.jpg",lbl_tmp )
#cv2.imwrite(save_name,lbl_tmp )
lbl_pil = PIL.Image.fromarray(lbl_tmp.astype(np.uint8))
lbl_pil.convert('1').save(save_name)
elif MODE == 'RGB':
lbl_pil = PIL.Image.fromarray(lbl.astype(np.uint8))
colormap = label_colormap(255)
lbl_pil.putpalette((colormap * 255).astype(np.uint8).flatten())
lbl_pil.convert('RGB').save(save_name)
else:
print('save mode error!')
#####################################################
for key, value in otherData.items():
data[key] = value
try:
with open(filename, 'wb' if PY2 else 'w') as f:
json.dump(data, f, ensure_ascii=False, indent=2)
self.filename = filename
except Exception as e:
raise LabelFileError(e)
@staticmethod
def is_label_file(filename):
return osp.splitext(filename)[1].lower() == LabelFile.suffix
- 修改完label_file.py， 重新运行labelme标注完成都既可以自动在保存的路径下创建mask文件夹，其中保存的是500x400 size的标注图像。因为我的项目的要求是resize到统一尺寸，也可以原图保存，只需要注释掉以上代码中的resize操作即可。
- 为了方便使用，也可以打包成exe可执行文件。
Setup conda
conda create --name labelme python==3.6.0
conda activate labelme
Build the standalone executable
pip install .
pip install pyinstaller
pyinstaller labelme.spec
dist/labelme --version