Multi-task learning of facial landmarks and attributes with Tensorflow（一）

数据集下载链接：

http://mmlab.ie.cuhk.edu.hk/projects/TCDCN.html

github地址：

https://github.com/Mimsyy/Deep-learning-MTFL 

Multi-Task Facial Landmark (MTFL人脸数据库)，这个数据库包括了12,995张人脸图片，每张图片都被做了一些标注。包括
（1）5个人脸特征点，包括左眼、右眼、鼻子、左嘴角、右嘴角。
（2）标记性别、微笑、眼镜、姿态。

标记数据格式

数据库被分成两个部分：training和tesing，标记的信息全部存放在txt文件中（traning和testing文件）。
文件中的每一行为一张人脸图片的标记信息，他的格式如下：
image path x1...x5,y1..y5 gender smile wearing glasses head pose
--x1...x5,y1...y5: 标记了左眼、右眼、鼻子、左嘴角、右嘴角的坐标位置。
--gender: 1代表男, 2代表女
--smile: 1代表微笑, 2代表不微笑。
--glasses: 1代表戴眼镜, 2代表没戴眼镜。
--head pose: 1 for left profile，2 for left, 3 for frontal, 4 for right, 5 for right profile

对数据预处理文件的解读

from PIL import Image
import os
import numpy as np
#数据集存放地址
folder = os.path.abspath(os.path.join("./", os.pardir)+"/MTFL") + "/"
#最终要将图全部resize成这个大小150*150
finalSize = 150
counter = 0
infoFiles = ["training.txt", "testing.txt"]
validation_counter = 0
#建立一个validation.txt文件用于存放验证数据信息
val_file = open(folder + "validation.txt","a")
#"training.txt", "testing.txt"中的信息进行处理
for idx in range(len(infoFiles)):
    info = infoFiles[idx]
    #当idx=0时读取training.txt，当idx=1时读取testing.txt
    f = open(folder+info,"r")
    #打开一个新的文件，用于写入信息
    fnew = open("tmp", 'a')
    #当读取training.txt信息时，创建aug_training.txt文件
    if idx == 0:
        fnew_augmented = open(folder+"aug_"+info,'a')
    lines = f.readlines()
    for line in lines:
        line = line.strip("\n ").split(" ")
        #获取到图片名称
        imgName = line[0].replace("\\", "/")
        if imgName == "":
            break
        #打开图片，读取图片的宽高
        img = Image.open(folder+imgName, mode='r')
        originalWidth = img.width
        originalHeight = img.height
        if (originalWidth != originalHeight):
            # Skip non-square images，若高宽不一致则跳过
            img.close()
            continue

        # Resize and check RGB，将图片resize到150*150
        if(originalWidth != finalSize):
            img = img.resize((finalSize, finalSize), Image.ANTIALIAS)    # Image.ANTIALIAS代表高质量

            # Check if img is RGB or greyscale，若图片不是彩色图，则跳过
            pixels = list(img.getdata())
            width, height = img.size
            pixelsArray = np.asarray([pixels[i * width:(i + 1) * width] for i in range(height)])
            if len(pixelsArray.shape) != 3:
                # img is greyscale, skip it
                img.close()
                continue

        coords = []
        # 图片的缩放比例，若大小没有被调整该系数为1，若图像大小调整，同比缩放坐标数据，记录在coords中line[i][1]-line[i][10]
        # 对应x1,x2,x3,x4,x5,y1,y2,y3,y4,y5: 标记了左眼、右眼、鼻子、左嘴角、右嘴角的坐标位置
        coordsScaleFactor = float(finalSize) / float(originalWidth)
        for i in range(1,11):
            coords.append(float(line[i])*coordsScaleFactor)
        #属性gender,smile,glasses,head pose分别对应每行line[i][11]-line[i][14]
        attributes = np.array([int(line[i]) for i in range(11,15)])

        #对testing.txt进行读取，生成1000个用于验证的图片信息于validation.txt中
        if(idx == 1 and validation_counter <= 1000):
            #line[i][0]
            val_file.write(imgName)
            #写入特征点信息
            for coord in coords:
                val_file.write(" " + str(coord))
            #写入属性信息
            for attribute in attributes:
                val_file.write(" " + str(attribute - 1)) # Subtract 1 for better indexing
            val_file.write("\n")
            validation_counter += 1
        # '''
        #         --gender: 1代表男, 2代表女
        #         --smile: 1代表微笑, 2代表不微笑。
        #         --glasses: 1代表戴眼镜, 2代表没戴眼镜。
        #         --head pose: 1 for left profile，2 for left, 3 for frontal, 4 for right, 5 for right profile
        #         attribute - 1:是为了使索引从0开始，方便训练
        # '''
        #剩下的写入fnew中用于训练
        else:
            # Write resized img to file
            fnew.write(imgName)
            for coord in coords:
                fnew.write(" " + str(coord))
            for attribute in attributes:
                fnew.write(" " + str(attribute - 1)) # Subtract 1 for better indexing
            fnew.write("\n")

        # Mirror the image if it's not part of test data
        if idx == 0:
            # Get the new img name,生成翻转图像的名称./MTFL/lfw_5590/Aaron_Eckhart_0001.jpg-->./MTFL/lfw_5590/Aaron_Eckhart_0001_transl.jpg
            splitName = imgName.split('.')
            imgNameTransp = splitName[0] + '_transl.' + splitName[1]

            # Mirror the image and save it
            #左右翻转图像，存图
            imgTransp = img.copy().transpose(Image.FLIP_LEFT_RIGHT)
            imgTransp.save(folder+imgNameTransp)
            imgTransp.close()
            #生成对应翻转图像特征点及属性特征
            coordsTransp = [0 for i in range(10)]
            # Translate x-coords for eyes, nose, and mouth
            coordsTransp[0] = 150 - coords[1]    #左眼睛x1
            coordsTransp[1] = 150 - coords[0]    #右眼睛x2
            coordsTransp[2] = 150 - coords[2]    #鼻子x3
            coordsTransp[3] = 150 - coords[4]    #左嘴角X4
            coordsTransp[4] = 150 - coords[3]    #右嘴角x5
            # Translate y-coords for eyes, nose, and mouth
            coordsTransp[5] = coords[6]          #左眼睛y1
            coordsTransp[6] = coords[5]          #右眼睛y2
            coordsTransp[7] = coords[7]          #鼻子y3
            coordsTransp[8] = coords[9]          #左嘴角y4
            coordsTransp[9] = coords[8]          #右嘴角y5
            # Translate attributes  属性变换gender,smile,glasses,head pose
            attributesTransp = np.array([int(line[i]) for i in range(11,15)])
            attributesTransp[3] = 6 - attributesTransp[3] # Translate head：1 for left profile，2 for left, 3 for frontal, 4 for right, 5 for right profile

            # Write resized old img to augmented file，写入原图的图像信息
            fnew_augmented.write(imgName)
            for coord in coords:
                fnew_augmented.write(" " + str(coord))
            for attribute in attributes:
                fnew_augmented.write(" " + str(attribute - 1))
            fnew_augmented.write("\n")

            # Write mirrored img to augmented file，写入翻转后的图像信息
            fnew_augmented.write(imgNameTransp)
            for coord in coordsTransp:
                fnew_augmented.write(" " + str(coord))
            for attribute in attributesTransp:
                fnew_augmented.write(" " + str(attribute - 1))
            fnew_augmented.write("\n")

        # Save resized img 存的都是resize到150*150的图片
        img.save(folder+imgName)
        img.close()
        
        counter = counter + 1
        if counter % 1000 == 0:
            print(counter,"files processed")
    #对"training.txt"/"testing.txt"中的每一行读完之后，就关闭该文件
    f.close()
    #对"training.txt"/"testing.txt"中的每一行数据进行操作完成，写入完毕后，就关闭该"tmp"文件
    fnew.close()
    #关闭aug_training.txt
    if idx == 0:
        fnew_augmented.close()
    #删除之前的"training.txt"/"testing.txt"
    os.remove(folder+info)
    #将tmp重命名为"training.txt"/"testing.txt"
    os.rename("tmp",folder+info)
#都循环完毕后，关闭validation.txt
val_file.close()