Please indicate the source https://blog.csdn.net/Fire_Light_/article/details/79589926
原文链接:Targeting Ultimate Accuracy: Face Recognition via Deep Embedding
Network architecture:
Are aligned good input face image, the face image by the face markers have Patch cut into the overlapping portion, respectively, of the input network between a plurality of mutually identical configuration, each network trained on different GPU .
9 convolutional network structure comprising layers with a layer behind some convolution cell layer, layer standardized paper are not described in detail in the figures, the network is generally know what it is.
When training with the training softmax, characterized in that all the FC verify contact layers, constituting a high-dimensional facial features.
Authentication method -Metric Learning measure learning:
By learning to triplet loss metric as the supervisory signal 128 to one kind of learning feature dimensions, the distance between the feature for authentication tasks.
The so-called measure learning after learning that some suitable distance measurement method in the feature space, it is equivalent to find a subspace of feature space, will feature transform into the sub-space, you can easily measure various characteristics the distance between (refer to "machine learning" - Zhou Zhihua).
By this measure learning, it can be done from within l2 reduced categories, the effect of increasing the distance l2 between classes.
Measure learning diagram:
a detailed description of the triplet loss can refer to my blog: Face series (six): FaceNet
experiment:
Affect the amount of training data error rate:
The amount of influence patch error rate
Optimal results:
7 using patches, wherein each of the final extract a 128-dimensional patch, and a number of other models are also mixed together is determined (in particular did not say what the text)
It reached 99.85% accuracy rate on LFW
The paper also showed all validation errors on the picture:
FIG divided into three categories:
. A error flag.
b. false negative, caused by make-up, glasses, characters relatively large changes or blocking other reasons.
c. a false positive, very similar facial features of people.