近期使用caffe模型进行人脸识别时，需要提取指定网络层中的数据进行特征比较，常见的比如提取全连接层等。本文基于caffe的python接口实现了一个简单的特征提取及人脸验证测试。测试使用的数据是LFW官网的6000对图片
本文主要参考caffe官网的例子： Image Classification and Filter Visualization
具体内容见代码注释: feature_extract_and_evaluate.py
# -*- coding: utf-8 -*- """ Created on Tue. Mar 07 16:59:55 2017 @author: wujiyang @brief：读取lfw数据库6000对人脸数据进行精确度验证 """ import numpy as np import matplotlib.pyplot as plt import os import sys import sklearn import sklearn.metrics.pairwise as pw caffe_root = '/home/wujiyang/caffe/' sys.path.insert(0, caffe_root + 'python') import caffe #图片库路径，若label.txt中使用的是绝对路径，则无需修改 imageBasePath = '' #模型初始化相关操作 def initilize(): print 'model initilizing...' deployPrototxt = "./vgg-face-deploy.prototxt" modelFile = "./vgg-face.caffemodel" caffe.set_mode_gpu() caffe.set_device(0) net = caffe.Net(deployPrototxt, modelFile,caffe.TEST) return net #读取图像标签 def readImagelist(labelFile): ''' @brief：从列表文件中，读取图像数据到矩阵文件中 @param： labelfile 图像列表文件 @return ：需要对比的图片路径，以及是否为同一个人的标记 labelfile的格式说明：每一行的数据如下所示，左边分别为要对比的图片，右边为是否是同一个人的标记 同一个表示为1，不同人表示为0 img1 img2 label ''' file = open(labelFile) lines = file.readlines() file.close() left = [] right = [] labels = [] for line in lines: path = line.strip('\n').split('\t') #read left image left.append(path[0]) #read right image right.append(path[1]) #read label labels.append(int(path[2])) assert(len(left) == len(right)) assert(len(right) == len(labels)) return left, right, labels #特征提取，使用caffe的io接口 def extractFeature(leftImageList, rightImageList, net): transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape}) # create transformer for the input called 'data' transformer.set_transpose('data', (2,0,1)) # move image channls to outermost dimension #transformer.set_mean('data', np.load(caffe_root + meanFile).mean(1).mean(1)) # mean pixel transformer.set_raw_scale('data', 255) # rescale from [0,1] to [0,255] transformer.set_channel_swap('data', (2,1,0)) # swap channels from RGB to BGR # set net to batch size of 1 如果图片较多就设置合适的batchsize net.blobs['data'].reshape(1, 3,224, 224) leftfeature = [] rightfeature = [] for i in range(len(leftImageList)): #分别读取左右列图像数据 #读取左边图像,并提取特征 imageleft = os.path.join(imageBasePath, leftImageList[i]) img = caffe.io.load_image(imageleft) #print imageleft, img.shape img = caffe.io.resize_image(img, (224,224)) #print imageleft, img.shape net.blobs['data'].data[...] = transformer.preprocess('data', img) #print net.blobs['data'].data.shape out = net.forward() #featureleft = np.float64(out['fc7']) #print featureleft.shape feature = np.float64(net.blobs['fc7'].data) #print feature, type(feature) leftfeature.append(feature) #读取右边图像,并提取特征 imageright = os.path.join(imageBasePath, rightImageList[i]) img = caffe.io.load_image(imageright) img = caffe.io.resize_image(img, (224,224)) net.blobs['data'].data[...] = transformer.preprocess('data', img) out = net.forward() feature = np.float64(net.blobs['fc7'].data) rightfeature.append(feature) return leftfeature, rightfeature #list #计算精确度 def calculate_accuracy(distance, labels, num): ''' #计算识别率, 选取阈值，计算识别率 ''' accuracy = {} predict = np.empty((num,)) threshold = 0.1 while threshold <= 0.9 : for i in range(num): if distance[i] >= threshold: predict[i] = 0 else: predict[i] = 1 predict_right =0.0 for i in range(num): if predict[i] == labels[i]: predict_right += 1.0 current_accuracy = (predict_right / num) accuracy[str(threshold)] = current_accuracy threshold = threshold + 0.001 #将字典按照value排序 temp = sorted(accuracy.items(), key = lambda d:d[1], reverse = True) highestAccuracy = temp[0][1] thres = temp[0][0] return highestAccuracy, thres #绘制roc曲线 def draw_roc_curve(fpr,tpr,title='cosine',save_name='roc_lfw'): #绘制roc曲线 plt.figure() plt.plot(fpr, tpr) plt.plot([0, 1], [0, 1], 'k--') plt.xlim([0.0, 1.0]) plt.ylim([0.0, 1.0]) plt.xlabel('False Positive Rate') plt.ylabel('True Positive Rate') plt.title('Receiver operating characteristic using: '+title) plt.legend(loc="lower right") plt.show() #plt.savefig(save_name+'.png') if __name__ == '__main__': #模型初始化 net = initilize() #读取文件列表 print "reading imagelist..." leftImageList, rightImageList, labels = readImagelist('tmp.txt') #提取特征 print "feature extracting..." leftfeature_list, rightfeature_list = extractFeature(leftImageList, rightImageList, net) #将特征list变为2维ndarray leftfeature = np.reshape(leftfeature_list,(len(labels),4096)) rightfeature = np.reshape(rightfeature_list,(len(labels),4096)) #print leftfeature.shape, rightfeature.shape #print leftfeature, type(leftfeature) #进行距离计算 cosine距离 print "computing distance..." dis = pw.pairwise_distances(leftfeature, rightfeature, metric='cosine') distance = np.empty((len(labels),)) for i in range(len(labels)): distance[i] = dis[i][i] print 'Distance before normalization:\n', distance print 'Distance max:', np.max(distance), 'Distance min:', np.min(distance), '\n' #距离需要归一化到0-1,与标签0-1匹配 distance_norm = np.empty((len(labels),)) for i in range(len(labels)): distance_norm[i] = (distance[i]-np.min(distance))/(np.max(distance)-np.min(distance)) #此处计算时会有误差产生，理论上应该有1的，然而却没有！ print 'Distance after normalization:\n', distance_norm print 'Distance_norm max:', np.max(distance_norm), 'Distance_norm min:', np.min(distance_norm), '\n' #根据label和distance_norm计算精确度 highestAccuracy, threshold = calculate_accuracy(distance_norm,labels,len(labels)) print ("The highest accuracy is : %.4f, and the corresponding threshold is %s \n"%(highestAccuracy, threshold)) #绘制roc曲线 fpr, tpr, thresholds = sklearn.metrics.roc_curve(labels, distance_norm) draw_roc_curve(fpr, tpr, title='cosine',save_name='lfw_evaluate')此脚本也可在我的github上找到 caffe-test-script
