做这件事首先需要在windows上搭建caffe环境,感谢 https://www.cnblogs.com/MY0213/p/9225310.html 提供思路
操作系统 windows 8.1 编译器 VS2015
第一步下载caffe的windows版本 https://github.com/BVLC/caffe/tree/windows
第二步下载cmake的windows安装包 https://blog.csdn.net/Create_IT_Man/article/details/105739974
第三步下载caffe的依赖包 (对应版本下)
WIN_DEPENDENCIES_URLS = {
('v120', '2.7'):("https://github.com/willyd/caffe-builder/releases/download/v1.1.0/libraries_v120_x64_py27_1.1.0.tar.bz2",
"ba833d86d19b162a04d68b09b06df5e0dad947d4"),
('v140', '2.7'):("https://github.com/willyd/caffe-builder/releases/download/v1.1.0/libraries_v140_x64_py27_1.1.0.tar.bz2",
"17eecb095bd3b0774a87a38624a77ce35e497cd2"),
('v140', '3.5'):("https://github.com/willyd/caffe-builder/releases/download/v1.1.0/libraries_v140_x64_py35_1.1.0.tar.bz2",
"f060403fd1a7448d866d27c0e5b7dced39c0a607"),
}
准备工作就绪,接着解压caffe在D盘,在D:\caffe-windows\scripts\build_win.cmd中,修改成如下,表示只是用cpu编译,使用python3.5编译pycaffe,python3.5的路径在D:\miniconda\envs\py35
接着用cmd运行build_win.cmd,确保cmd能用cmake。build_win.cmd运行到一半会去下依赖包,直接给ctrl+c中断,把下载好的依赖库压缩包放在C:\Users\(你的名字)\.caffe\dependencies\download\libraries_v140_x64_py35_1.1.0.tar.bz2,在没运行build_win.cmd之前.caffe这个文件夹是没有的。运行完的效果:
然后需要编译caffe,先将依赖库压缩包解压至caffe-windows\build下,如图
并将
D:\caffe-windows\build\libraries\bin、D:\caffe-windows\build\libraries\lib、D:\caffe-windows\build\libraries\x64\vc14\bin 三个路径添加至环境变量,接着使用VS2015打开D:\caffe-windows\scripts\build\Caffe.sln,直接生成解决方案,没有错误的话警告没关系。
最后,将D:\caffe-windows\python\caffe复制到你的python路径下即可。如图:
测试python+caffe(如果缺skimage就pip install)
---------------------------------------------------------------------------------------------------------------------------------------
人脸识别,本质上计算两个人脸的相似度
先准备VGGFACE的人脸识别模型 http://www.robots.ox.ac.uk/~vgg/software/vgg_face/
准备三张图片
import caffe
import numpy as np
import cv2 as cv
import sklearn.metrics.pairwise as pw
def get_feature(net, img):
img = cv.resize(img, (224,224))
img = cv.cvtColor(img, cv.COLOR_BGR2RGB).astype(np.float32)/255.0
img = img[:,:,::-1]*255.0
mean = np.array([129.1863,104.7624,93.5940])
img = img - mean
img = img.transpose((2,0,1))
img = img[None,:]
output = net.forward_all( data = img )
output_prob = output['fc7']
return output_prob.reshape(1, 4096)
caffe.set_mode_cpu()
proto_path = 'VGG_FACE_deploy.prototxt'
model_path = 'VGG_FACE.caffemodel'
net = caffe.Net(proto_path, model_path, caffe.TEST)
img1 = cv.imread('a.jpg')
img2 = cv.imread('b1.jpg')
img3 = cv.imread('b2.jpg')
f1 = get_feature(net, img1)
f2 = get_feature(net, img2)
f3 = get_feature(net, img3)
s1 = pw.cosine_similarity(f1,f2)
s2 = pw.cosine_similarity(f2,f3)
print('Similarity: %s' % s1)
print('Similarity: %s' % s2)
得到输出