Please indicate the source https://blog.csdn.net/Fire_Light_/article/details/79594590
论文链接：Recover Canonical-View Faces in the wild with Deep
Neural Networks
FR: means the face recovery FCN refers to the face convolution network
Frontal face reconstruction:
This is a 14-year paper, more interesting is the author uses a simple convolution neural network to complete the reconstruction of the front face, to align face, then face verification based on alignment of the human face.
To reconstruct the front face, a human face to recover, first of all related to the frontal face a choice as a matter of GroudTruth, that is our goal reconstruction, and therefore need to focus on selected choice we feel right frontal face from training .
Frontal face the choice:
- Bilateral symmetry, 2. rank Rank image , 3, binding (Method articles use) 1 and 2.
Therefore, a measure applying the following formula:
Wherein Yi is a human face image, P, Q matrix is a parameter,
The first term symmetry,
The second term of nuclear matrix norm: singular value matrix and the image can be approximated rank.
λ indicates tradeoff of these two criteria.
In the article, the author simply uses the smallest M value. (May be a problem, or that there is room for improvement, paper, the authors said, and can be calculated using a linear combination of the front face)
After the frontal face selection, you can train a network of deep learning, the training of its loss function is:
Where W is the depth of the parameters of the neural network, Yi is the choice of front face.
Network structure is as follows:
Convolution which comprises three layers, wherein the first two using max pooling, the last all-connected, not shared weights , particularly with classic network structure is not much improved.
Network architecture:
- For each of the training images, the reconstruction front face, and then extracted five feature points Landmark, and then extract the Patch based on these Landmark.
- With each of the patch to train the network, the plurality of features are finally concatenated together to form the final feature.
3. Network Logistic regression was used as a tail as a loss function, according to the first input to predict whether two pictures are of the same class, considering that this is an article earlier recognition, face recognition that time is not now generally level , so there is no better use of Softmax is understandable.
LFW score:
Get 96.45% accuracy rate on LFW