Multi-task Convolutional Neural Network 训练人脸识别和关键点检测 （python+tensoflow实现代码 算法步骤）---数据集处理_多数据集多任务学习人脸识别

1.数据处理

将用于PFLD的数据集处理成只有嘴巴包围盒+标注的数据集

首先通过matplotlib库函数画出所有的标记，在通过删减点的个数来确定关键点标记在嘴巴的位置

例如使用IBUG人脸标记数据集做测试，每张图片有106个标记点位置，其中第84-104 有20个标记点是嘴巴

# 将数据集处理成嘴巴标记和嘴巴包围盒的数据
 
import cv2
import os
import  matplotlib
 
#将人脸图片画上标记输出展示效果
picture_set = []
 
#读取原图片目录
picture_path = "./traindata/IBUG/picture"  #原图片目录
for item in os.listdir(picture_path):
    if ('jpg' not in item):
        continue
    picture_set.append(os.path.join(picture_path, item))
 
 
save_path = "./output_landmark"    #输出图像目录
if not os.path.exists(save_path):
         os.makedirs(save_path)
 
image = cv2.imread(picture_set[0])
land_path = "./traindata/IBUG/landmark"
land_path = os.path.join(land_path, (picture_set[0] + ".txt").split("/")[-1])
print(land_path)
count = 0
# 33 个面部轮廓点 < 35
# 最后两个关键点是眼睛的中心 106 107 去掉
# 85 -105 行 是嘴巴关键点的标注
with open(land_path) as read_file:
     for line in read_file:
         count = count + 1
         if(count < 85) :continue
         pos = line.split()
         x=int(float(pos[0])*1000000)
         y=int(float(pos[1])*1000000)
         x1=int(float(pos[0]))
         x2=int(float(pos[1]))
         #画出标记
         print(count)
         cv2.circle(image,(x1,x2),2,(0,0,255),-1)
 
cv2.imwrite(save_path + "/%s.jpg" % (land_path.split("/")[-1]), image)
 
 
 
 
 
 

所以将每个数据集的关键点位置文本标记只留下第84-104行 的坐标位置，并且计算出包围盒数据另外输出在一个txt文件中

--输出嘴巴关键点标记：

# 将数据集处理成嘴巴标记和嘴巴包围盒的数据
 
import cv2
import os
 
 
#将人脸图片画上标记输出展示效果
picture_set = []
 
dirpath="IBUG"  #图片所在文件夹名字
 
#读取原图片目录
picture_path = "./traindata/"+dirpath+"/picture"  #原图片目录
for item in os.listdir(picture_path):
    if ('jpg' not in item):
        continue
    picture_set.append(os.path.join(picture_path, item))
 
 
save_path = "./output_landmark"    #输出图像目录
if not os.path.exists(save_path):
         os.makedirs(save_path)
 
 
pic_num=0
for imagepath in picture_set:
     count = 0
     land_path = "./traindata/"+dirpath+"/landmark"
     land_path = os.path.join(land_path, (imagepath + ".txt").split("/")[-1])
     with open(save_path + "/%s" % (land_path.split("/")[-1]),'w') as input_file:  #打开输入文件
         with open(land_path) as read_file:
             for line in read_file:
                 count = count + 1
                 if (count < 85 or count > 105): continue
                 input_file.write(line)
                 # 画出标记
     pic_num=pic_num+1
     print("%d done" % pic_num)
 
# 33 个面部轮廓点 < 35
# 最后两个关键点是眼睛的中心 106 107 去掉
# 85 -105 行 是嘴巴关键点的标注

----计算嘴巴包围盒数据并保存在另一个文件夹box中

# 将数据集处理成嘴巴标记和嘴巴包围盒的数据
 
import cv2
import os
 
save_path = "./box"    #输出图像目录
if not os.path.exists(save_path):
         os.makedirs(save_path)
 
source_set=[] #源目录文件列表
source_path="./landmark"
for item in os.listdir(source_path):
    if ('txt' not in item):
        continue
    source_set.append(os.path.join(source_path, item))
 
test=[]
test.append(source_set[0])
 
pic_num=0
for path in source_set:
     count = 0
     with open(save_path + "/%s" % (path.split("/")[-1]),'w') as input_file:  #打开输入文件
         max_x = 0.00
         max_y = 0.00
         min_x = 9999.9
         min_y = 9999.9
         with open(path) as read_file:
             for line in read_file:
                 pos = line.split()
                 x = float(pos[0])
                 y = float(pos[1])
                 if(x>max_x):max_x=x
                 if(x<min_x):min_x=x
                 if(y>max_y):max_y=y
                 if(y<min_y):min_y=y
         box="%f %f %f %f" %(min_x,min_y,max_x-min_x,max_y-min_y)
         input_file.write(box)
 
                 # 画出标记
     pic_num=pic_num+1
     print("%d done" % pic_num)
 
#测试包围盒是否准确
# image = cv2.imread("./traindata/IBUG/picture/IBUG_image_003_1_0.jpg")
# 
# with open("./box/IBUG_image_003_1_0.jpg.txt") as read_file:
#      for line in read_file:
#          pos = line.split()
#          x1=int(float(pos[0]))
#          x2=int(float(pos[1]))
#          width=int(float(pos[2]))
#          height=int(float(pos[3]))
#          #画出标记
#          print("%d %d %d %d" %(x1,x2,width,height))
#          cv2.rectangle(image,(x1,x2),(x1+width,x2+height),(0,0,255),3)
# 
# 
# cv2.imwrite("out.jpg", image)
 
 
 

///***整合包围盒数据，标记点数据，格式化数据

import os
 
 
source_set=[]
source_path = "./box"
 
pic_path="./picture/"
land_path="./landmark/"
 
for item in os.listdir(source_path):
    if ('txt' not in item):
        continue
    source_set.append(os.path.join(source_path, item))
 
with open("mouth_train.txt",'a') as input:  #写入包围盒数据
    for path in source_set:
        with open(path) as elem:
            content=elem.readline()
            filename= path.split("/")[-1]
            filename=filename.strip(".txt")
            input.write(pic_path+filename+" ")
            input.write(content+"\n")
 
with open("mouth_train_box.txt",'a') as input:  #整合包围合数据
    for path in source_set:
        with open(path) as elem:
            content=elem.readline()
            filename= path.split("/")[-1]
            filename=filename.strip(".txt")
            input.write(pic_path+filename+"\n")
            input.write("1\n")
            input.write(content+"\n")
 
 
with open("trainImageList.txt",'a') as input:  #整合关键点数据
    for path in source_set:
        with open(path) as elem:
            content=elem.readline()
            filename= path.split("/")[-1]
            filename=filename.strip(".txt")
            input.write(pic_path+filename+" ")
            input.write(content+" ")
 
            filename = path.split("/")[-1]
            filename=land_path + filename
            with open(filename) as f:
                for line in f:
                    line=line.strip('\n')
                    input.write(line)
            input.write("\n")