Point to set similarity comparison and deep feature learning for visual recognition
A visual recognition system to process images includes a global sub-network including a convolutional layer and a first max pooling layer. A local sub-network is connected to receive data from the global sub-network, and includes at least two convolutional layers, each connected to a max pooling layer. A fusion network is connected to receive data from the local sub-network, and includes a plurality of fully connected layers that respectively determine local feature maps derived from images. A loss layer is connected to receive data from the fusion network, set filter parameters, and minimize ranking error.
1. A visual recognition system, comprising:
a global sub-network including a convolutional layer and a first max pooling layer;
a local sub-network connected to receive data from the global sub-network, and including at least two convolutional layers, each connected to a max pooling layer;
a fusion network connected to receive data from the local sub-network, and including a plurality of fully connected layers that respectively determine local feature maps derived from images; and
a loss layer connected to receive data from the fusion network and set filter parameters, the loss layer implementing a symmetric triplet loss function configured to: (a) both maximize inter-class distances and minimize intra-class distances by fine tuning directions of gradient back-propagation and (b) train the visual recognition system to learn filter weights and biases that minimize ranking error.
2. The visual recognition system of claim 1 , wherein the symmetric triplet loss function is further configured to improve ranking accuracy by jointly keeping similarity of positive image pairs and dissimilarity of negative image pairs.
3. The visual recognition system of claim 1 , wherein fine tuning directions of gradient back-propagation comprise optimizing parameters for a similarity metric that includes a pairwise term and a triplet term.
4. The visual recognition system of claim 1 , wherein the symmetric triplet loss function is implemented using
X i (m) =Ψ( W (m) *X i (m−1) +b (m) ), i =1, 2, . . . , N; m =1, 2, . . . , M; X i (0) =X i .
5. The visual recognition system of claim 1 , wherein the fusion layer further comprises rectified linear units (ReLU) connected between fully connected layers.
6. The visual recognition system of claim 1 , wherein the dimension of the fully connected layers in the fusion layer is at least 100.
7. The visual recognition system of claim 1 , wherein the dimension of the fully concatenated connected layers representing image featured in the fusion layer is at least 800.
8. The visual recognition system of claim 1 , wherein the at least two convolutional layers of the local sub-network includes a plurality of filters.
9. The visual recognition system of claim 1 , wherein the global sub-network further comprises rectified linear units (ReLU).
10. A visual recognition image processing method comprising:
providing a global sub-network including a convolutional layer and a first max pooling layer to extract features from images and form feature maps;
passing the feature maps to a local sub-network connected to the global sub-network to discriminately learn features;
generating final feature representations using a fusion network connected to receive data from the local sub-network; and
using a loss layer connected to receive data from the fusion network and set filter parameters, the loss layer implementing a symmetric triplet loss function configured to: (a) both maximize inter-class distances and minimize intra-class distances by fine tuning directions of gradient back-propagation and (b) train the visual recognition system to learn filter weights and biases that minimize ranking error.
11. The visual recognition system of claim 1 , wherein the symmetric triplet loss function comprises a multi-layer model recursive function that integrates the filter weights, the biases, a convolution operation, an element-wise non-linear activation function, and the feature maps.
12. The visual recognition image processing method of claim 10 , wherein the symmetric triplet loss function is further configured to improve ranking accuracy by jointly keeping similarity of positive image pairs and dissimilarity of negative image pairs.
13. The visual recognition image processing method of claim 10 , wherein fine tuning directions of gradient back-propagation comprise optimizing parameters for a similarity metric that includes a pairwise term and a triplet term.
14. The visual recognition image processing method of claim 10 , wherein the symmetric triplet loss function comprises a multi-layer model recursive function that integrates the filter weights, the biases, a convolution operation, an element-wise non-linear activation function, and the feature maps.
15. The visual recognition image processing method of claim 10 , wherein the symmetric triplet loss function is implemented using
X i (m) =Ψ( W (m) *X i (m−1) +b (m) ), i =1, 2, . . . , N; m =1, 2, . . . , M; X i (0) =X i .
16. A system comprising:
a processor; and
system memory coupled to the processor and storing instructions configured to cause the processor to:
provide a global sub-network including a convolutional layer and a first max pooling layer to extract features from images and form feature maps;
pass the feature maps to a local sub-network connected to the global sub-network to discriminately learn features;
generate final feature representations using a fusion network connected to receive data from the local sub-network; and
use a loss layer connected to receive data from the fusion network and set filter parameters, the symmetric loss function implementing a symmetric triplet loss function configured to: (a) both maximize inter-class distances and minimize intra-class distances by fine tuning directions of gradient back-propagation and (b) train the visual recognition system to learn filter weights and biases that minimize ranking error.
17. The system of claim 16 , wherein the symmetric triplet loss function is further configured to improve ranking accuracy by jointly keeping similarity of positive image pairs and dissimilarity of negative image pairs.
18. The system of claim 16 , wherein fine tuning directions of gradient back-propagation comprise optimizing parameters for a similarity metric that includes a pairwise term and a triplet term.
19. The system of claim 16 , wherein the symmetric triplet loss function comprises a multi-layer model recursive function that integrates the filter weights, the biases, a convolution operation, an element-wise non-linear activation function, and the feature maps.
20. The system of claim 16 , wherein the symmetric triplet loss function is implemented using
X i (m) =Ψ( W (m) *X i (m−1) +b (m) ), i =1, 2, . . . , N; m =1, 2, . . . , M; X i (0) =X i .