分割數據集label轉換為目標檢測boundingbox

實現功能

將分割的label圖轉換為目標檢測boundingbox標注文件（VOC格式）。

注：

1.分割樣本里一張圖片只有同一類別的多個目標。

2.轉換為boundingbox標注通過連通域實現，所以重疊的目標處理不了，會標為1個。

數據集格式

其中，語義分割數據集格式如下：

原圖片在JPEGImages文件夾中，命名格式為ImageID.jpg

　　　　

Label圖在labelimage文件夾中，命名格式為ImageID_classname.png

　　　　　

生成的boundingbox標注命名格式為ImageID.xml

　　　

XML標注格式

<annotation>
   <folder>road_dataset</folder>                      #文件名
   <filename>3425.jpg</filename>                      #原圖片名
   <path>D:\road_dataset\JPEGImages\3425.jpg</path>   #原圖片地址
   <source>
      <database>Unknown</database>
   </source>
   <size>                               #圖片尺寸
      <width>512</width>
      <height>512</height>
      <depth>3</depth>
   </size>
   <segmented>0</segmented>            #是否用於分割，0為否
   <object>                            #目標
      <name>butt</name>                #類別名稱
      <pose>Unspecified</pose>         #拍攝角度
      <truncated>0</truncated>         #是否被截斷
      <difficult>0</difficult>         #是否為困難樣本
      <bndbox>                         #boundingbox坐標（左下、右上）
         <xmin>327</xmin>
         <ymin>38</ymin>
         <xmax>394</xmax>
         <ymax>69</ymax>
      </bndbox>
   </object>
   <object>                             #多個目標
      <name>Cigarette butts</name>
      <pose>Unspecified</pose>
      <truncated>0</truncated>
      <difficult>0</difficult>
      <bndbox>
         <xmin>139</xmin>
         <ymin>279</ymin>
         <xmax>214</xmax>
         <ymax>318</ymax>
      </bndbox>
   </object>
</annotation>

其中<pose>  <truncated> <difficult> 全是默認值。

得到label圖中的連通域

使用skimage的morphology, measure通過連通域得到每一副一幅圖片上的目標數量和boundingbox。

import os
import numpy as np
from itertools import groupby
from skimage import morphology,measure
from PIL import Image
from scipy import misc

# 因為一張圖片里只有一種類別的目標，所以label圖標記只有黑白兩色
rgbmask = np.array([[0,0,0],[255,255,255]],dtype=np.uint8)

# 從label圖得到 boundingbox 和圖上連通域數量 object_num
def getboundingbox(image):
    # mask.shape = [image.shape[0], image.shape[1], classnum]
    mask = np.zeros((image.shape[0], image.shape[1]), dtype=np.uint8)
    mask[np.where(np.all(image == rgbmask[1],axis=-1))[:2]] = 1
    # 刪掉小於10像素的目標
    mask_without_small = morphology.remove_small_objects(mask,min_size=10,connectivity=2)
    # 連通域標記
    label_image = measure.label(mask_without_small)
    #統計object個數
    object_num = len(measure.regionprops(label_image))
    boundingbox = list()
    for region in measure.regionprops(label_image):  # 循環得到每一個連通域bbox
        boundingbox.append(region.bbox)
    return object_num, boundingbox

在label圖片上顯示boundingbox，查看結果：

import matplotlib.pyplot as plt
import matplotlib.patches as patch

# 輸出成圖片查看得到boundingbox效果
imagedir = r'D:\test_dataset\labelimage'

if ~os.path.exists(r'D:\test_dataset\test_getbbox'):
    os.mkdir(r'D:\test_dataset\test_getbbox')
for root, _, fnames in sorted(os.walk(imagedir)):
    for fname in sorted(fnames):
        imagepath = os.path.join(root, fname)
        image = misc.imread(imagepath)
        objectnum, bbox = getboundingbox(image)
        ImageID = fname.split('.')[0]
        
        fig,ax = plt.subplots(1)
        ax.imshow(image)
        for box in bbox:
            rect = patch.Rectangle((box[1], box[0]), box[3]-box[1], box[2]-box[0],edgecolor = 'r', linewidth = 1,fill = False)
            ax.add_patch(rect)
        plt.savefig('D:/test_dataset/test_getbbox/'+ImageID+'.png')

輸出圖像為：

生成XML標注文件

createXMLlabel： 根據標注信息生成XML標注文件

import xml.etree.ElementTree as ET

def createXMLlabel(savedir,objectnum, bbox, classname, foldername='0',filename='0', path='0', database='road', width='400', height='600',depth='3', segmented='0', pose="Unspecified", truncated='0', difficult='0'):
    # 創建根節點
    root = ET.Element("annotation")

    # 創建子節點
    folder_node = ET.Element("folder")
    folder_node.text = foldername
    # 將子節點數據添加到根節點
    root.append(folder_node)

    file_node = ET.Element("filename")
    file_node.text = filename
    root.append(file_node)
    path_node = ET.Element("path")
    path_node.text = path
    root.append(path_node)

    source_node = ET.Element("source")
    # 也可以使用SubElement直接添加子節點
    db_node = ET.SubElement(source_node, "database")
    db_node.text = database
    root.append(source_node)

    size_node = ET.Element("size")
    width_node = ET.SubElement(size_node, "width")
    height_node = ET.SubElement(size_node, "height")
    depth_node = ET.SubElement(size_node, "depth")
    width_node.text = width
    height_node.text = height
    depth_node.text = depth
    root.append(size_node)

    seg_node = ET.Element("segmented")
    seg_node.text = segmented
    root.append(seg_node)

    for i in range(objectnum):
        newEle = ET.Element("object")
        name = ET.Element("name")
        name.text = classname
        newEle.append(name)
        pose_node = ET.Element("pose")
        pose_node.text = pose
        newEle.append(pose_node)
        trunc = ET.Element("truncated")
        trunc.text = truncated
        newEle.append(trunc)
        dif = ET.Element("difficult")
        dif.text = difficult
        newEle.append(dif)
        boundingbox = ET.Element("bndbox")
        xmin = ET.SubElement(boundingbox, "xmin")
        ymin = ET.SubElement(boundingbox, "ymin")
        xmax = ET.SubElement(boundingbox, "xmax")
        ymax = ET.SubElement(boundingbox, "ymax")
        xmin.text = str(bbox[i][1])
        ymin.text = str(bbox[i][0])
        xmax.text = str(bbox[i][3])
        ymax.text = str(bbox[i][2])
        newEle.append(boundingbox)
        root.append(newEle)

    ImageID = filename.split('.')[0]
    # 創建elementtree對象，寫入文件
    tree = ET.ElementTree(root)
    tree.write(savedir + '/'+ ImageID + ".xml")

 

imagedir = r'D:\test_dataset\labelimage'
saveXMLdir = r'D:\test_dataset\Annotations'

if os.path.exists(saveXMLdir) is False:
    os.mkdir(saveXMLdir)

for root, _, fnames in sorted(os.walk(imagedir)):
    for fname in sorted(fnames):
        labelpath = os.path.join(root, fname)
        labelimage = misc.imread(labelpath)
        # 得到label圖上的boundingingbox和數量
        objectnum, bbox = getboundingbox(labelimage)
        # label圖 命名格式為 ImgeID_classname.png
        labelfilename = labelpath.split('\\')[-1]
        ImageID = labelfilename.split('.')[0].split('_')[0]
        classname = labelfilename.split('.')[0].split('_')[1]
        origin_image_name = ImageID +'.jpg'
    
        # 一些圖片信息
        foldername = 'test_dataset'
        path  ='\\'.join(imagedir.split('\\')[:-1]) + '\\JPEGImage\\'+ origin_image_name
        database = 'Unknown'
        width = str(labelimage.shape[0])
        height = str(labelimage.shape[1])
        depth = str(labelimage.shape[2])
        
        createXMLlabel(saveXMLdir,objectnum, bbox, classname, foldername=foldername,filename=origin_image_name, path=path,
                       database=database, width=width, height=height,depth=depth, segmented='0', pose="Unspecified",
                       truncated='0', difficult='0')