实现功能
将分割的label图转换为目标检测boundingbox标注文件（VOC格式）。
注：
1.分割样本里一张图片只有同一类别的多个目标。
2.转换为boundingbox标注通过连通域实现，所以重叠的目标处理不了，会标为1个。
数据集格式
其中，语义分割数据集格式如下：
原图片在JPEGImages文件夹中，命名格式为ImageID.jpg
Label图在labelimage文件夹中，命名格式为ImageID_classname.png
生成的boundingbox标注命名格式为ImageID.xml
XML标注格式
<annotation><folder>road_dataset</folder> #文件名<filename>3425.jpg</filename> #原图片名<path>D:\road_dataset\JPEGImages\3425.jpg</path> #原图片地址<source><database>Unknown</database></source><size> #图片尺寸<width>512</width><height>512</height><depth>3</depth></size><segmented>0</segmented> #是否用于分割，0为否<object> #目标<name>butt</name>#类别名称<pose>Unspecified</pose> #拍摄角度<truncated>0</truncated> #是否被截断<difficult>0</difficult> #是否为困难样本<bndbox>#boundingbox坐标（左下、右上）<xmin>327</xmin><ymin>38</ymin><xmax>394</xmax><ymax>69</ymax></bndbox></object><object> #多个目标<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> 全是默认值。
得到label图中的连通域
使用skimage的morphology, measure通过连通域得到每一副一幅图片上的目标数量和boundingbox。
import osimport numpy as npfrom itertools import groupbyfrom skimage import morphology,measurefrom PIL import Imagefrom scipy import misc# 因为一张图片里只有一种类别的目标，所以label图标记只有黑白两色rgbmask = np.array([[0,0,0],[255,255,255]],dtype=np.uint8)# 从label图得到 boundingbox 和图上连通域数量 object_numdef 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# 删掉小于10像素的目标mask_without_small = morphology.remove_small_objects(mask,min_size=10,connectivity=2)# 连通域标记label_image = measure.label(mask_without_small)#统计object个数object_num = len(measure.regionprops(label_image))boundingbox = list()for region in measure.regionprops(label_image): # 循环得到每一个连通域bboxboundingbox.append(region.bbox)return object_num, boundingbox
在label图片上显示boundingbox，查看结果：
import matplotlib.pyplot as pltimport matplotlib.patches as patch# 输出成图片查看得到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')
输出图像为：
生成XML标注文件
createXMLlabel： 根据标注信息生成XML标注文件
import xml.etree.ElementTree as ETdef 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'):# 创建根节点root = ET.Element("annotation")# 创建子节点folder_node = ET.Element("folder")folder_node.text = foldername# 将子节点数据添加到根节点root.append(folder_node)file_node = ET.Element("filename")file_node.text = filenameroot.append(file_node)path_node = ET.Element("path")path_node.text = pathroot.append(path_node)source_node = ET.Element("source")# 也可以使用SubElement直接添加子节点db_node = ET.SubElement(source_node, "database")db_node.text = databaseroot.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 = widthheight_node.text = heightdepth_node.text = depthroot.append(size_node)seg_node = ET.Element("segmented")seg_node.text = segmentedroot.append(seg_node)for i in range(objectnum):newEle = ET.Element("object")name = ET.Element("name")name.text = classnamenewEle.append(name)pose_node = ET.Element("pose")pose_node.text = posenewEle.append(pose_node)trunc = ET.Element("truncated")trunc.text = truncatednewEle.append(trunc)dif = ET.Element("difficult")dif.text = difficultnewEle.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]# 创建elementtree对象，写入文件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图上的boundingingbox和数量objectnum, bbox = getboundingbox(labelimage)# label图 命名格式为 ImgeID_classname.pnglabelfilename = labelpath.split('\\')[-1]ImageID = labelfilename.split('.')[0].split('_')[0]classname = labelfilename.split('.')[0].split('_')[1]origin_image_name = ImageID +'.jpg'# 一些图片信息foldername = 'test_dataset'path ='\\'.join(imagedir.split('\\')[:-1]) + '\\JPEGImage\\'+ origin_image_namedatabase = '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')
