如何更高效的标注文字检测算法的label(二)

在做自然场景下的文字检测算法时，第一步就是要标注文字位置的label，如果手动从头开始标注就太累了，所以我们小组总结出一种方法，试了试比较高效，因此分享出来。

先使用一个基础的baseline算法，然后把真实的样本过一遍baseline的模型，每张样本图片会生成相应的bounddingbox信息的txt文件，接着使用脚本将这些txt文件反转成标注软件能识别的xml格式(或者json)，然后将这些xml格式的文件导入到标注软件，打开标注软件的时候，就可以看到bounddingbox在图片上的编辑位置，最后就可以对这些方框进行编辑修改，正确的可以忽略，错误的进行调整。

对于标注软件，我们使用过标注精灵助手和IphotoDraw两款软件，最后还是觉得IphotoDraw功能比较强大和方便，所以最后选择了IphotoDraw作为标注软件，下面是将baseline模型跑的boundingbox位置点信息txt文件，转成IphotoDraw能识别的xml格式文件：

代码如下：

# -*- coding: utf-8 -*-
"""
Created on Fri Feb  2 09:57:25 2018@author: new
"""
import os
import math
import numpy as np# 将txt中所有四边形角点生成对应xml文件，进行标注def endWith(s,*endstring):array=map(s.endswith,endstring)if True in array:return Trueelse:return Falsedef writeXml(xmlName,txtName,imageName):xml_file=open(xmlName,'w',encoding='utf-8')xml_file.write('<!--Document-->\n')xml_file.write('<Document FileVersion="1.0">\n')xml_file.write('    <ExportImageSettings FileName='+'"'+imageName+'"'+'/>\n')xml_file.write('    <Layers>\n')xml_file.write('    <Layer Name="Layer1" Visible="True" LockedShapesIndex="">\n')xml_file.write('        <Shapes>\n')fopen=open(txtName,encoding='utf-8')for line in fopen.readlines():# txtData = line.split(',')# print(txtData)p,w,h,angle,label = get_coords(line)if w > 0 and h > 0:num = 3text = labelxml_file.write('<Shape Type="Rectangle">\n')xml_file.write('<Settings>\n')xml_file.write('<MiscSettings GroupRendering="Unknown" />\n')xml_file.write('<Font Name="Arial" Size="4" Style="Regular">\n')xml_file.write('<Color Alpha="255" R="0" G="0" B="0" />\n')xml_file.write('</Font>\n')xml_file.write('<Line Width="1" Dash="Solid" Join="Round" OutlineType="Color" DashOffset="False" StartRoundCap="False" EndRoundCap="False">\n')if int(float(num)) % 3 == 0:xml_file.write('<Color Alpha="255" R="26" G="170" B="66" />\n')elif int(float(num)) % 3 == 1:xml_file.write('<Color Alpha="255" R="34" G="72" B="234" />\n')elif int(float(num)) % 3 == 2:xml_file.write('<Color Alpha="255" R="255" G="0" B="0" />\n')xml_file.write('</Line>\n')xml_file.write('<Fill FillType="None">\n')# if int(float(num)) % 3 == 0:#     xml_file.write('<Color Alpha="58" R="26" G="170" B="66" />\n')# elif int(float(num)) % 3 == 1:#     xml_file.write('<Color Alpha="58" R="34" G="72" B="234" />\n')# elif int(float(num)) % 3 == 2:#     xml_file.write('<Color Alpha="58" R="234" G="22" B="30" />\n')xml_file.write('<Color Alpha="255" R="255" G="255" B="255" />\n')xml_file.write('<GradientSettings Type="Linear" Angle="0" HorizontalOffset="0" VerticalOffset="0" StartExtension="0" EndExtension="0" BoundaryResize="100">\n')xml_file.write('<StartingColor Alpha="255" R="0" G="0" B="0" />\n')xml_file.write('<EndingColor Alpha="255" R="255" G="255" B="255" />\n')xml_file.write('<Blend />')xml_file.write('</GradientSettings>')xml_file.write('<EmbeddedImage Align="Center" ImageFillType="Stretch" Alpha="255" FileName="">\n')xml_file.write('<StretchSettings Type="KeepOriginalSize" Align="Center" ZoomFactor="100">\n')xml_file.write('<Offset X="0" Y="0" />\n')xml_file.write('</StretchSettings>\n')xml_file.write('<TileSettings WrapMode="Tile">\n')xml_file.write('<Offset X="0" Y="0" />\n')xml_file.write('</TileSettings>\n')xml_file.write('<ImageOptions Rotation="0">\n')xml_file.write('<Flip HorizontalFlip="False" VerticalFlip="False" />\n')xml_file.write('</ImageOptions>\n')xml_file.write('<ImageData><![CDATA[]]></ImageData>\n')xml_file.write('</EmbeddedImage>\n')xml_file.write('</Fill>\n')xml_file.write('<TextEffect UseTextEffect="False" />\n')xml_file.write('<EffectSettings>\n')xml_file.write('<Shadow UseShadow="False" Angle="45" Offset="5" Size="100" BlurLevel="0">\n')xml_file.write('<Color Alpha="255" R="0" G="0" B="0" />\n')xml_file.write('</Shadow>\n')xml_file.write('<Glow UseGlow="False" BlurLevel="20" Thickness="8">\n')xml_file.write('<Color Alpha="255" R="29" G="199" B="244" />\n')xml_file.write('</Glow>\n')xml_file.write('<WavyLine UseWavyLine="False" WavePattern="CosineSmooth" Ridges="5" Height="20" VerticalFlip="False" OffsetAtStartPoint="0" OffsetAtEndPoint="0" />\n')xml_file.write('</EffectSettings>\n')xml_file.write('</Settings>\n')xml_file.write('<BlockText Align="Center" VerticalAlign="Middle" RightToLeft="Unknown">\n')xml_file.write('<Text>' + text + '</Text>\n')xml_file.write('<Margin Left="0" Top="0" Right="0" Bottom="0" />\n')xml_file.write('</BlockText>\n')xml_file.write('<Data IsRoundCorner="False" RoundCornerRadius="0" Rotation="' + str(angle) +'">\n')xml_file.write('<Extent X=')xml_file.write('"' + str(p[0]) + '"')xml_file.write(' Y=')xml_file.write('"' + str(p[1]) + '"')xml_file.write(' Width=')xml_file.write('"' + str(w) + '"')xml_file.write(' Height=')xml_file.write('"' + str(h) + '"')xml_file.write('/>\n')xml_file.write('</Data>\n')xml_file.write('</Shape>\n')xml_file.write('</Shapes>\n')xml_file.write('</Layer>\n')xml_file.write(' </Layers>\n')xml_file.write('<Snapshots />\n')xml_file.write('</Document>')xml_file.close()fopen.close()def get_new_coord(center_coord,ori_coord,rotate_angle):x_new = (ori_coord[0]-center_coord[0])*math.cos((rotate_angle/180.)*math.pi)+(ori_coord[1]-center_coord[1])*math.sin((rotate_angle/180.)*math.pi)+center_coord[0]y_new = (ori_coord[1]-center_coord[1])*math.cos((rotate_angle/180.)*math.pi)-(ori_coord[0]-center_coord[0])*math.sin((rotate_angle/180.)*math.pi)+center_coord[1]return x_new,y_newdef get_rotation_coord(iphotodraw_result):result = []center_X = 1/2*(iphotodraw_result[0]+iphotodraw_result[0]+iphotodraw_result[2])center_Y = 1/2*(iphotodraw_result[1]+iphotodraw_result[1]+iphotodraw_result[3])x1,y1 = iphotodraw_result[0],iphotodraw_result[1]x2,y2 = iphotodraw_result[0]+iphotodraw_result[2],iphotodraw_result[1]x3,y3 = iphotodraw_result[0]+iphotodraw_result[2],iphotodraw_result[1]+iphotodraw_result[3]x4,y4 = iphotodraw_result[0],iphotodraw_result[1]+iphotodraw_result[3]result.extend(cal_coord((center_X,center_Y),(x1,y1),iphotodraw_result[-1]))result.extend(cal_coord((center_X,center_Y),(x2,y2),iphotodraw_result[-1]))result.extend(cal_coord((center_X,center_Y),(x3,y3),iphotodraw_result[-1]))result.extend(cal_coord((center_X,center_Y),(x4,y4),iphotodraw_result[-1]))return resultdef cal_coord(center_coord,ori_coord,angle):angle = angle*math.pi/180out_x = math.cos(angle)*(ori_coord[0]-center_coord[0])-math.sin(angle)*(ori_coord[1]-center_coord[1])+center_coord[0]out_y = math.sin(angle)*(ori_coord[0]-center_coord[0])+math.cos(angle)*(ori_coord[1]-center_coord[1])+center_coord[1]return [out_x,out_y]def coord_to_iphotodrawFormat(bbox):### bbox shape 1*8angle = math.atan((bbox[3]-bbox[1])/(bbox[2]-bbox[0]))*(180/math.pi)width = math.sqrt((bbox[3]-bbox[1])**2+(bbox[2]-bbox[0])**2)height = math.sqrt((bbox[5]-bbox[3])**2+(bbox[4]-bbox[2])**2)center_coord = [1 / 2 * (bbox[0] + bbox[4]), 1 / 2 * (bbox[1] + bbox[5])]ori_coord = [bbox[0], bbox[1]]ori_coord = cal_coord(center_coord, ori_coord,-math.atan((bbox[3] - bbox[1]) / (bbox[2] - bbox[0])) * 180 / math.pi)return (ori_coord[0],ori_coord[1]),width,height,angledef get_coords(line):"""文本检测结果:param txt_path: CTPN结果路径:return:"""try:label = line.strip().split(',')[-1]point1_x = int(line.strip().split(',')[0]) #右下角点,顺时针point1_y = int(line.strip().split(',')[1])point1 = [point1_x,point1_y]point2_x = int(line.strip().split(',')[2])point2_y = int(line.strip().split(',')[3])point2 = [point2_x, point2_y]point3_x = int(line.strip().split(',')[4])point3_y = int(line.strip().split(',')[5])point3 = [point3_x, point3_y]point4_x = int(line.strip().split(',')[6])point4_y = int(line.strip().split(',')[7])point4 = [point4_x, point4_y]bbox = []bbox.extend(point1)bbox.extend(point2)bbox.extend(point3)bbox.extend(point4)p,w,h,angle = coord_to_iphotodrawFormat(bbox)return p, w, h, angle, labelexcept:return (0,0), 0, 0, 0, '#'def sortPoint(points,center):'''将四边形的四个角点进行排序，返回 左上角，左下角，右下角，右上角:param point_list::return:'''idx_list = np.where(np.array(points)[:, 0] > center[0])[0]right_point = [points[idx] for idx in idx_list]right_point = sorted(right_point, key=lambda x: x[1])right_up = right_point[0]right_bottom = right_point[1]idx_list = np.where(np.array(points)[:, 0] < center[0])[0]left_point = [points[idx] for idx in idx_list]left_point = sorted(left_point, key=lambda x: x[1])left_up = left_point[0]left_bottom = left_point[1]return left_up,left_bottom,right_bottom,right_upimport cv2if __name__=='__main__':#生成的xml在basepath文件夹下basepath = r'xxx'out_path = r'xxx'import shutilfrom PIL import Imageif not os.path.exists(out_path):os.makedirs(out_path)# jpgNames = os.listdir(basepath)# for name in jpgNames:#     if endWith(name,'.txt'):##         xmlName = name.split('.')[0].split('_')[0] + '_data.xml'#         imageName = name.split('.')[0].split('_')[0] + '.jpg'##         if os.path.exists(os.path.join(basepath,imageName)):#             xmlName = os.path.join(basepath,xmlName)#             txtName = os.path.join(basepath,name)#             writeXml(xmlName,txtName,imageName)jpgNames = os.listdir(basepath)for name in jpgNames:if endWith(name, '.txt'):print(name)xmlName = name.replace('.txt', '_data.xml')# IMIMkey = name.replace('_gt_for_xml.txt','').split('_')[-1]imageName = name.replace('.txt', '.jpg')if os.path.exists(os.path.join(basepath, imageName)):xmlName = os.path.join(out_path, xmlName)txtName = os.path.join(basepath, name)# image = Image.open(os.path.join(basepath, imageName))# image = image.convert('RGB')# image = np.array(image)image = cv2.imread(os.path.join(basepath, imageName))cv2.imwrite(os.path.join(out_path, imageName),image)# shutil.copy( imageName,os.path.join(out_path,name.replace('_gt_for_xml.txt', '.jpg') ))writeXml(xmlName, txtName, imageName)

如何更高效的标注文字检测算法的label(二)相关推荐

基于深度学习的目标检测算法综述（二）
转自:https://zhuanlan.zhihu.com/p/40020809 基于深度学习的目标检测算法综述(一) 基于深度学习的目标检测算法综述(二) 基于深度学习的目标检测算法综述(三) 本文 ...
【文字检测算法整理】
文字检测与其他目标检测的区别: 一.长宽比差异很大,而且普遍较小: 二.文字是以字符为基本单元按照一定空间排列而成的序列,而不是一个单独的目标: 三.文字存在多种粒度和多语言. 传统方法系列: 一.流 ...
SlimYOLOv3：更窄、更快、更好的无人机目标检测算法
点击我爱计算机视觉标星,更快获取CVML新技术无人机因为硬件计算能力较弱,要在其上实现实时的目标检测,需要算法参数量小.占用内存少.推断时间短.常见的算法往往难以直接应用. 一种比较直接的做法是对模 ...
前景检测算法（十二）--基于模糊Choquet积分
原文: http://www.cnblogs.com/pangblog/p/3303956.html 本文根据论文:Fuzzy Integral for Moving Object Detection ...
ACM MM'21 | 超轻量8.5M！更高效的RGB-D显著性检测模型DFM-Net
点击上方"3D视觉工坊",选择"星标" 干货第一时间送达自深度传感器的普及以来,RGB-D显著物体检测(Salient object detection,SO ...
场景文本检测算法可微分二值化DBNet原理与代码解析
目录原理介绍 Label Generation Loss函数后处理论文 https://arxiv.org/abs/1911.08947 代码 https://github.com/WenmuZ ...
自然场景下的文字检测：从多方向迈向任意形状
点击我爱计算机视觉标星,更快获取CVML新技术本文经作者授权转载自知乎旷视Detection组专栏: https://zhuanlan.zhihu.com/p/68058851 旷视检测组在刚刚结束 ...
【项目实践】中英文文字检测与识别项目（CTPN+CRNN+CTC Loss原理讲解）
点击上方"小白学视觉",选择加"星标"或"置顶" 重磅干货,第一时间送达本文转自:opencv学堂 OCR--简介文字识别也是图像领域一 ...
大盘点｜YOLO 系目标检测算法总览
点击上方"3D视觉工坊",选择"星标" 干货第一时间送达 YOLO目标检测算法诞生于2015年6月,从出生的那一天起就是"高精度.高效率.高实用性&q ...

如何更高效的标注文字检测算法的label(二)

如何更高效的标注文字检测算法的label(二)相关推荐

最新文章

热门文章