一、区别

  1. 语义分割:每个像素都打上标签(这个像素是人,树,背景等)语义分割只区分类别,不区分类别中的具体单元

实例分割:不光要区分类别,还要区分类别中的每一个个体

全景分割:相当于语义分割加上实例分割

二、代码演示原图像与mask融合

from PIL import Image

import imgviz

import numpy as np

image_file=r'D:\aaa\envs\labelme\Scripts\2_3_json\img.png'

mask_file=r'D:\aaa\envs\labelme\Scripts\2_3_json\label.png'

image=Image.open(image_file)

mask=Image.open(mask_file)

mask_img=Image.blend(image.convert("RGBA"),
                     mask.convert("RGBA"),0.5)

mask_img.save("vis2.png")

三、数据处理

将labelme标注好的json转换成mask图像

import json
import os

import imgviz
import numpy as np
from PIL import Image
import cv2
import glob

def save_colored_mask(mask,image_file):
    lbl_image=Image.fromarray(mask.astype(np.uint8),mode='P')
    colormap=imgviz.label_colormap()
    lbl_image.putpalette(colormap.flatten())
    lbl_image.save(image_file)


json_files=r'E:\桌面\资料\语义分割'
img_file=r'E:\桌面\资料\语义分割\图片'

json_l=glob.glob(os.path.join(json_files,'*.json'))

for json_ in json_l:
    name=os.path.basename(json_)
    img_name=name.replace('json','png')
    fs=open(json_,encoding='utf-8')
    dict_=json.load(fs)
    # 获取图像 宽,高
    height = dict_['imageHeight']
    width = dict_['imageWidth']

    shapes = dict_["shapes"]
    # 生成一个全零图像
    img = np.zeros((height, width), dtype=np.uint8)

    label_color = {"sheep": 1}
    for shape in shapes:
        # 解析多边形轮廓点的坐标
        points = shape['points']
        # 解析多边形的标签
        label = shape['label']

        points = np.array(points, dtype=np.int32)
        # 绘制轮廓
        cv2.polylines(img, [points], isClosed=True, color=(255), thickness=2)
        # 填充多边形颜色
        cv2.fillPoly(img, [points], color=label_color[label])
    img_path=os.path.join(img_file,img_name)
    print(img_path)

    save_colored_mask(img, img_path)