Receptive field learning for pooled image features
View Patent ↗Systems and methods are disclosed for image classification by receiving an overcomplete set of spatial regions, jointly optimizing the classifier and the pooling region for each pooled feature; and performing incremental feature selection and retraining using a grafting process to efficiently train the classifier.
1. An image classification method, comprising:
receiving an overcomplete set of spatial regions;
jointly optimizing the classifier and the pooling region for each pooled feature;
performing incremental feature selection and retraining using a grafting process to efficiently train the classifier, and
training a multiple-class linear classifier y=f(x)=Wx+b with an optimization problem:
min
W
,
b
1
N
∑
n
=
1
N
l
(
W
T
x
n
+
b
,
y
n
)
+
λ
1
1
W
Fro
2
+
λ
2
W
1
,
∞
where Y i is the L-dimensional label vector coded in a 1-of-L fashion, with values taken from {−1,+1} given L classes x i is an M-dimensional feature vector defined by overcomplete pooling in the previous subsection, and W=[w 1 , w 2 , . . . , w L ] is a M×L weight matrix containing the weight vector for the L classifiers.
2. The method of claim 1 , comprising extracting local image patches and converting the images to a set of overcomplete representation (coding).
3. The method of claim 1 , comprising determining statistics of local representations over predefined spatial regions (pooling).
4. The method of claim 1 , comprising collecting statistics to perform a final image classification task.
5. The method of claim 1 , comprising optimizing with a multi-class extension of the binomial negative log likelihood (BNLL) loss function:
l
(
W
T
x
+
b
,
y
)
=
∑
i
=
1
L
ln
(
1
+
ⅇ
-
y
i
(
w
.
,
i
T
x
+
b
i
)
)
.
6. The method of claim 1 , comprising performing fast approximate learning.
7. The method of claim 6 , for each feature index j, comprising determining a score of a feature as a 2-norm of a gradient of an objective function, denoted by L(W,b), with respect to weight vectors:
score
(
j
)
=
∂
L
(
W
,
b
)
∂
W
j
Fro
2
.
8. The method of claim 7 , comprising approximately retraining a model on selected subset.
9. The method of claim 1 , comprising performing local patch based, multiple-layer pipelinng to find predetermined image features for a task.
10. An image classification system, comprising:
means forreceiving an overcomplete set ofspatial regions, and means for jointly optimizing the classifier and the pooling region for each pooled feature;
means for performing incremental feature selection and retraining using a grafting processto efficiently train the classifier, and
means for training a multiple-class linear classifier y=f(x)=Wx+b with an optimization problem:
min
W
,
b
1
N
∑
n
=
1
N
l
(
W
T
x
n
+
b
,
y
n
)
+
λ
1
1
W
Fro
2
+
λ
2
W
1
,
∞
where Y i is the L-dimensional label vector coded in a 1-of-L fashion with values taken from {−1,+1} given L classes, X i is an M-dimensional feature vector defined by overcomplete pooling in the previous subsection, and W=[w 1 , w 2 , . . . , w L ] is a M×L weight matrix containing the weight vector for the L classifiers.
11. The system of claim 10 , comprising means for determining statistics of local representations over predefined spatial regions (pooling).
12. The system of claim 10 , comprising means for collecting statistics to perform a final image classification task.
13. The system of claim 10 , comprising means for optimizing with a multi-class extension of the binomial negative log likelihood (BNLL) loss function:
l
(
W
T
x
+
b
,
y
)
=
∑
i
=
1
L
ln
(
1
+
ⅇ
-
y
i
(
w
.
,
i
T
x
+
b
i
)
)
.
14. The system of claim 10 , comprising means for performing fast approximate learning.
15. The system of claim 14 , for each feature index j, comprising determining a score of a feature as a 2-norm of a gradient of an objective function, denoted by L(W,b) with respect to weight vectors:
score
(
j
)
=
∂
L
(
W
,
b
)
∂
W
j
Fro
2
.
16. The system of claim 15 , comprising approximately retraining model on selected subset.
17. The system of claim 10 , comprising a local patch based, multiple-layer pipeline to find predetermined image features for a task.
18. An image classification method, comprising:
receiving an overcomplete set of spatial regions;
jointly optimizing the classifier and the pooling region for each pooled feature;
performing incremental feature selection and retraining using a grafting process to efficiently train the classifier; and
performing fast approximate learning;
wherein for each feature index j, comprising determining a score of a feature as a 2-norm of a gradient of an objective function, denoted by L(W,b), with respect to weight vectors:
score
(
j
)
=
∂
L
(
W
,
b
)
∂
W
j
Fro
2
.