facenet lfw训练

10 篇文章 1 订阅

https://blog.csdn.net/oYeZhou/article/details/88942598

近期,做人脸识别项目,用到了facenet这个开源框架,并使用LFW人脸数据集进行了测试。现将该过程总结如下:
1 facenet简介

GitHub地址:https://github.com/davidsandberg/facenet.git

facenet的原理就是基于同一人脸总是比不同人脸更相似这一先验知识,然后利用传统卷积神经网络特征提取,利用三元损失函数进行训练。最终,将人脸映射到特征空间后,同一身份的人脸距离较近,不同身份的人脸距离较远。模型的输出是一个512维的向量(原来是128维)。

算法详情可参考其论文:https://arxiv.org/pdf/1503.03832.pdf。
2 LFW数据集简介

网盘链接: https://pan.baidu.com/s/1qOrFv_8RhIhUJvAmwE8p0g 提取码: kfwh

LFW数据集是对5000多人在自然场景下采集的共13000多张图像。lfw_funneled文件夹中每个子文件夹代表一个人,其中包含其若干张同一身份不同场景下的照片,有的只有一张,有的有多张。

lfw_funneled中还包含了几个txt文档,这里面记录了这些人脸的不同组合,我们使用其中的pairs.txt中的组合进行人脸比对测试。

pairs.txt里面包含了6000对人脸,3000对同一身份,3000对不同身份。文档第一行的10  300代表正负样本以300的数量依次罗列,重复10次,因此共10*(300对正样本+300对负样本)= 6000对人脸。

3  测试过程
3.1 图像路径提取

首先,我们根据pairs.txt进行图片路径的提取:

    def get_img_pairs_list(pairs_txt_path,img_path):
        """ 指定图片组合及其所在文件,返回各图片对的绝对路径
            Args:
                pairs_txt_path:图片pairs文件,里面是6000对图片名字的组合
                img_path:图片所在文件夹
            return:
                img_pairs_list:深度为2的list,每一个二级list存放的是一对图片的绝对路径
        """
        file = open(pairs_txt_path)
        img_pairs_list,labels = [],[]
        while 1:
            img_pairs = []
            line = file.readline().replace('\n','')
            if line == '':
                break
            line_list = line.split('\t')
            if len(line_list) == 3:
                # 图片路径示例:
                # 'C:\Users\thinkpad1\Desktop\image_set\lfw_funneled\Tina_Fey\Tina_Fey_0001.jpg'
                img_pairs.append(img_path+'\\'+line_list[0]+'\\'+line_list[0]+'_'+('000'+line_list[1])[-4:]+'.jpg')
                img_pairs.append(img_path+'\\'+line_list[0]+'\\'+line_list[0]+'_'+('000'+line_list[2])[-4:]+'.jpg')
                labels.append(1)
            elif len(line_list) == 4:
                img_pairs.append(img_path+'\\'+line_list[0]+'\\'+line_list[0]+'_'+('000'+line_list[1])[-4:]+'.jpg')
                img_pairs.append(img_path+'\\'+line_list[2]+'\\'+line_list[2]+'_'+('000'+line_list[3])[-4:]+'.jpg')
                labels.append(0)
            else:
                continue
            
            img_pairs_list.append(img_pairs)
        return img_pairs_list,labels

利用上述代码,即可提取所有人类对的绝对路径,返回一个路径list及其标签(1或0)。
3.2 人脸检测、对比

获取到人脸对的图片路径及标签之后,在使用facenet将其转化为512维的向量之前,需要先对图像进行人脸提取,即截取其中的人脸区域。这里用到了MTCNN模型,用于检测出人脸并将人脸区域单独提出来,然后就可以利用facenet进行人脸特征向量的转化了。得到这对人脸的特征向量之后,求其欧氏距离,即可根据该距离判断其是否为同一身份了。提取及比对过程如下(其中模型model是MTCNN的参数,在facenet的GitHub项目的“facenet/src/models/”路径下已有;model_facenet模型因为比较大,需要单独下载,点击下载:链接: https://pan.baidu.com/s/1ty7NfBYIretHhnZwwl2dTg 提取码: g3jy):

    def face_verification(img_pairs_list):
        model = './model/'
        model_facenet = r'XXX\XXX\20180402-114759.pb' # 模型在你电脑中的路径
        # mtcnn相关参数
        minsize=40
        threshold=[0.4,0.5,0.6] # pnet、rnet、onet三个网络输出人脸的阈值,大于阈值则保留,小于阈值则丢弃
        factor = 0.709  # scale factor
        
        # 创建mtcnn网络
        with tf.Graph().as_default():
            sess=tf.Session()
            with sess.as_default():
                pnet,rnet,onet=detect_face.create_mtcnn(sess, model)
        
        margin = 44
        image_size = 160
        
        with tf.Graph().as_default():
            
            with tf.Session() as sess:
                
                # 根据模型文件载入模型
                facenet.load_model(model_facenet)
                # 得到输入、输出等张量
                images_placeholder = tf.get_default_graph().get_tensor_by_name("input:0")
                embeddings = tf.get_default_graph().get_tensor_by_name("embeddings:0")
                phase_train_placeholder = tf.get_default_graph().get_tensor_by_name("phase_train:0")
                
                # 设置可视化进度条相关参数
                jd = '\r   %2d%%\t [%s%s]'
                bar_num_total = 50    
                total_num = len(img_pairs_list)
                result, dist = [],[]
                
                for i in range(len(img_pairs_list)):
                    
                    # 画进度条
                    if i%round(total_num/bar_num_total) == 0 or i == total_num-1:
                        bar_num_alright = round(bar_num_total*i/total_num)
                        alright = '#'*bar_num_alright
                        not_alright = '□'*(bar_num_total-bar_num_alright)
                        percent = (bar_num_alright/bar_num_total)*100
                        print(jd % (percent,alright,not_alright),end='')
                    
                    # 读取一对人脸图像
                    img_pairs = img_pairs_list[i]
                    img_list = []
                    img1 = cv2.imread(img_pairs[0])
                    img2 = cv2.imread(img_pairs[1])
                    
                    img_size1 = np.asarray(img1.shape)[0:2]
                    img_size2 = np.asarray(img2.shape)[0:2]
                    
                    # 检测该对图像中的人脸
                    bounding_box1,_1=detect_face.detect_face(img1,minsize,pnet,rnet,onet,threshold,factor)
                    bounding_box2,_2=detect_face.detect_face(img2,minsize,pnet,rnet,onet,threshold,factor)
                    
                    # 未检测到人脸,则将结果标为-1,后续计算准确率时排除
                    if len(bounding_box1)<1 or len(bounding_box2)<1:
                        result.append(-1)
                        dist.append(-1)
                        continue
                    
                    # 将图片1加入img_list
                    det = np.squeeze(bounding_box1[0,0:4])
                    bb = np.zeros(4, dtype=np.int32)
                    bb[0] = np.maximum(det[0]-margin/2, 0)
                    bb[1] = np.maximum(det[1]-margin/2, 0)
                    bb[2] = np.minimum(det[2]+margin/2, img_size1[1])
                    bb[3] = np.minimum(det[3]+margin/2, img_size1[0])
                    cropped = img1[bb[1]:bb[3],bb[0]:bb[2],:]
                    aligned = cv2.resize(cropped, (image_size, image_size))
                    prewhitened = facenet.prewhiten(aligned)
                    img_list.append(prewhitened)
                    
                    # 将图片2加入img_list
                    det = np.squeeze(bounding_box2[0,0:4])
                    bb = np.zeros(4, dtype=np.int32)
                    bb[0] = np.maximum(det[0]-margin/2, 0)
                    bb[1] = np.maximum(det[1]-margin/2, 0)
                    bb[2] = np.minimum(det[2]+margin/2, img_size2[1])
                    bb[3] = np.minimum(det[3]+margin/2, img_size2[0])
                    cropped = img2[bb[1]:bb[3],bb[0]:bb[2],:]
                    aligned = cv2.resize(cropped, (image_size, image_size))
                    prewhitened = facenet.prewhiten(aligned)
                    img_list.append(prewhitened)
                    
                    images = np.stack(img_list)
                    
                    # 将两个人脸转化为512维的向量
                    feed_dict = { images_placeholder: images, phase_train_placeholder:False }
                    emb = sess.run(embeddings, feed_dict=feed_dict)
                    
                    # 计算两个人脸向量的距离
                    ed = np.sqrt( np.sum( np.square( np.subtract(emb[0], emb[1]) ) ) )
                    dist.append(ed)
                    # 根据得出的人脸间的距离,判断是否属于同一个人
                    if ed<=1.1:
                        result.append(1)
                    else:
                        result.append(0)
        return result,dist

上述代码可以实现在某一指定阈值下,进行人脸比对,得出对比结果存于result中,用于后续计算准确率;同时,为了画出ROC曲线,这里还返回了,所有人脸对的欧氏距离,存于dist中。

实际上,上述result是dist在某一个阈值下的截面数据,通过设置不同阈值,即可根据dist得出不同的result,下面正是利用这个原理画出的ROC曲线。
3.3 ROC曲线

根据3.2得出的每对人脸的欧氏距离,还有3.1得出的各对人脸样本的标签,即可画出计算出ROC曲线所需指标:TPR、FPR。

代码如下:

    def roc(dist,labels):
        TP_list,TN_list,FP_list,FN_list,TPR,FPR = [],[],[],[],[],[]
        for t in range(180):
            threh = 0.1+t*0.01
     
            TP,TN,FP,FN = 0,0,0,0
            for i in range(len(dist)):
                if labels[i]==1 and dist[i]!=-1:
                    if dist[i]<threh:
                        TP += 1
                    else:
                        FN += 1
                elif labels[i]==0 and dist[i]!=-1:
                    if dist[i]>=threh:
                        TN += 1
                    else:
                        FP += 1
            TP_list.append(TP)
            TN_list.append(TN)
            FP_list.append(FP)
            FN_list.append(FN)
            TPR.append(TP/(TP+FN))
            FPR.append(FP/(FP+TN))
        return TP_list,TN_list,FP_list,FN_list,TPR,FPR

 4 完整代码

    # -*- coding: utf-8 -*-
    """
    Created on Fri Mar 22 09:59:41 2019
    @author: Leon
    内容:
    人脸验证准确率测试
    样本:LFW人脸集,共6000对人脸,中3000对同一身份、3000对不同身份。
    """
    import numpy as np
    import cv2
    import tensorflow as tf
    import matplotlib.pyplot as plt
    # facenet 和 detect_face 均在facenet项目文件中,这里是直接将其放到测试脚本同一路径下了,也可以安装facenet,然后调用之
    import facenet
    import align.detect_face as detect_face
     
    def face_verification(img_pairs_list):
        model = './model/'
        model_facenet = r'XXX\XXX\20180402-114759.pb'
        # mtcnn相关参数
        minsize=40
        threshold=[0.4,0.5,0.6] # pnet、rnet、onet三个网络输出人脸的阈值,大于阈值则保留,小于阈值则丢弃
        factor = 0.709  # scale factor
        
        # 创建mtcnn网络
        with tf.Graph().as_default():
            sess=tf.Session()
            with sess.as_default():
                pnet,rnet,onet=detect_face.create_mtcnn(sess, model)
        
        margin = 44
        image_size = 160
        
        with tf.Graph().as_default():
            
            with tf.Session() as sess:
                
                # 根据模型文件载入模型
                facenet.load_model(model_facenet)
                # 得到输入、输出等张量
                images_placeholder = tf.get_default_graph().get_tensor_by_name("input:0")
                embeddings = tf.get_default_graph().get_tensor_by_name("embeddings:0")
                phase_train_placeholder = tf.get_default_graph().get_tensor_by_name("phase_train:0")
                
                # 设置可视化进度条相关参数
                jd = '\r   %2d%%\t [%s%s]'
                bar_num_total = 50    
                total_num = len(img_pairs_list)
                result, dist = [],[]
                
                for i in range(len(img_pairs_list)):
                    
                    # 画进度条
                    if i%round(total_num/bar_num_total) == 0 or i == total_num-1:
                        bar_num_alright = round(bar_num_total*i/total_num)
                        alright = '#'*bar_num_alright
                        not_alright = '□'*(bar_num_total-bar_num_alright)
                        percent = (bar_num_alright/bar_num_total)*100
                        print(jd % (percent,alright,not_alright),end='')
                    
                    # 读取一对人脸图像
                    img_pairs = img_pairs_list[i]
                    img_list = []
                    img1 = cv2.imread(img_pairs[0])
                    img2 = cv2.imread(img_pairs[1])
                    
                    img_size1 = np.asarray(img1.shape)[0:2]
                    img_size2 = np.asarray(img2.shape)[0:2]
                    
                    # 检测该对图像中的人脸
                    bounding_box1,_1=detect_face.detect_face(img1,minsize,pnet,rnet,onet,threshold,factor)
                    bounding_box2,_2=detect_face.detect_face(img2,minsize,pnet,rnet,onet,threshold,factor)
                    
                    # 未检测到人脸,则将结果标为-1,后续计算准确率时排除
                    if len(bounding_box1)<1 or len(bounding_box2)<1:
                        result.append(-1)
                        dist.append(-1)
                        continue
                    
                    # 将图片1加入img_list
                    det = np.squeeze(bounding_box1[0,0:4])
                    bb = np.zeros(4, dtype=np.int32)
                    bb[0] = np.maximum(det[0]-margin/2, 0)
                    bb[1] = np.maximum(det[1]-margin/2, 0)
                    bb[2] = np.minimum(det[2]+margin/2, img_size1[1])
                    bb[3] = np.minimum(det[3]+margin/2, img_size1[0])
                    cropped = img1[bb[1]:bb[3],bb[0]:bb[2],:]
                    aligned = cv2.resize(cropped, (image_size, image_size))
                    prewhitened = facenet.prewhiten(aligned)
                    img_list.append(prewhitened)
                    
                    # 将图片2加入img_list
                    det = np.squeeze(bounding_box2[0,0:4])
                    bb = np.zeros(4, dtype=np.int32)
                    bb[0] = np.maximum(det[0]-margin/2, 0)
                    bb[1] = np.maximum(det[1]-margin/2, 0)
                    bb[2] = np.minimum(det[2]+margin/2, img_size2[1])
                    bb[3] = np.minimum(det[3]+margin/2, img_size2[0])
                    cropped = img2[bb[1]:bb[3],bb[0]:bb[2],:]
                    aligned = cv2.resize(cropped, (image_size, image_size))
                    prewhitened = facenet.prewhiten(aligned)
                    img_list.append(prewhitened)
                    
                    images = np.stack(img_list)
                    
                    # 将两个人脸转化为512维的向量
                    feed_dict = { images_placeholder: images, phase_train_placeholder:False }
                    emb = sess.run(embeddings, feed_dict=feed_dict)
                    
                    # 计算两个人脸向量的距离
                    ed = np.sqrt( np.sum( np.square( np.subtract(emb[0], emb[1]) ) ) )
                    dist.append(ed)
                    # 根据得出的人脸间的距离,判断是否属于同一个人
                    if ed<=1.1:
                        result.append(1)
                    else:
                        result.append(0)
        return result,dist
     
    def get_img_pairs_list(pairs_txt_path,img_path):
        """ 指定图片组合及其所在文件,返回各图片对的绝对路径
            Args:
                pairs_txt_path:图片pairs文件,里面是6000对图片名字的组合
                img_path:图片所在文件夹
            return:
                img_pairs_list:深度为2的list,每一个二级list存放的是一对图片的绝对路径
        """
        file = open(pairs_txt_path)
        img_pairs_list,labels = [],[]
        while 1:
            img_pairs = []
            line = file.readline().replace('\n','')
            if line == '':
                break
            line_list = line.split('\t')
            if len(line_list) == 3:
                # 图片路径示例:
                # 'C:\Users\thinkpad1\Desktop\image_set\lfw_funneled\Tina_Fey\Tina_Fey_0001.jpg'
                img_pairs.append(img_path+'\\'+line_list[0]+'\\'+line_list[0]+'_'+('000'+line_list[1])[-4:]+'.jpg')
                img_pairs.append(img_path+'\\'+line_list[0]+'\\'+line_list[0]+'_'+('000'+line_list[2])[-4:]+'.jpg')
                labels.append(1)
            elif len(line_list) == 4:
                img_pairs.append(img_path+'\\'+line_list[0]+'\\'+line_list[0]+'_'+('000'+line_list[1])[-4:]+'.jpg')
                img_pairs.append(img_path+'\\'+line_list[2]+'\\'+line_list[2]+'_'+('000'+line_list[3])[-4:]+'.jpg')
                labels.append(0)
            else:
                continue
            
            img_pairs_list.append(img_pairs)
        return img_pairs_list,labels
     
    def roc(dist,labels):
        TP_list,TN_list,FP_list,FN_list,TPR,FPR = [],[],[],[],[],[]
        for t in range(180):
            threh = 0.1+t*0.01
     
            TP,TN,FP,FN = 0,0,0,0
            for i in range(len(dist)):
                if labels[i]==1 and dist[i]!=-1:
                    if dist[i]<threh:
                        TP += 1
                    else:
                        FN += 1
                elif labels[i]==0 and dist[i]!=-1:
                    if dist[i]>=threh:
                        TN += 1
                    else:
                        FP += 1
            TP_list.append(TP)
            TN_list.append(TN)
            FP_list.append(FP)
            FN_list.append(FN)
            TPR.append(TP/(TP+FN))
            FPR.append(FP/(FP+TN))
        return TP_list,TN_list,FP_list,FN_list,TPR,FPR
     
    if __name__ == '__main__':
        pairs_txt_path = 'C:/Users/thinkpad1/Desktop/image_set/lfw_funneled/pairs.txt'
        img_path = 'C:/Users/thinkpad1/Desktop/image_set/lfw_funneled'
        img_pairs_list,labels = get_img_pairs_list(pairs_txt_path,img_path)
        
        result,dist = face_verification(img_pairs_list)
        
        num_right, num_total = 0, 0
        num_total = len([r for r in result if r != -1])
        num_right = len([result[i] for i in range(len(result)) if result[i] == labels[i]])
        
        print("人脸验证测试完毕")
        print("阈值为1.1,共%d对人脸,准确率%2.4f%%"%(num_total, round(100*num_right/num_total,4)))
        
        TP_list,TN_list,FP_list,FN_list,TPR,FPR = roc(dist,labels)
        plt.plot(FPR,TPR,label='Roc')
        plt.plot([0, 1], [0, 1], '--', color=(0.6, 0.6, 0.6), label='Luck')
        plt.xlabel('FPR')
        plt.ylabel('TPR')
        plt.legend()
        
        plt.plot(np.linspace(0.1,1.89,180),TP_list,label='TP')
        plt.plot(np.linspace(0.1,1.89,180),TN_list,label='TN')
        plt.plot(np.linspace(0.1,1.89,180),FP_list,label='FP')
        plt.plot(np.linspace(0.1,1.89,180),FN_list,label='FN')
        plt.legend()

5 测试结果

在阈值1.1下测试准确率为93.48%,这里没有达到其宣称的99%+的准确率。

利用每对人脸距离,通过设置不同距离阈值,画出ROC曲线,如下图(左),将TP,TN,FP,FN的曲线也画出来,可以佐证阈值在1.1时,达到最好的分类效果(TP、TN最大,FP、FN最小)。

 
https://blog.csdn.net/oYeZhou/article/details/88942598

  • 0
    点赞
  • 13
    收藏
    觉得还不错? 一键收藏
  • 0
    评论

“相关推荐”对你有帮助么?

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值