Anomaly detection for deep neural networks
An image including a first object can be input to a deep neural network trained to detect objects. The deep neural network can output a first feature vector corresponding to the first object. A first distance can be measured from the first feature vector to a feature vector subspace determined using a k-means single value decomposition algorithm on an overcomplete dictionary of feature vectors. The first object can be determined to correspond to an anomaly based on the first distance.
1. A computer, comprising:
a processor; and
a memory, the memory including instructions executable by the processor to:
input an image including a first object to a deep neural network trained to detect objects;
output a first feature vector corresponding to the first object from the deep neural network;
measure a first distance from the first feature vector to a feature vector subspace determined using a K-means single value decomposition algorithm on an overcomplete dictionary of feature vectors; and
determine that the first object is an anomaly based on the first distance being greater than a bound determined based on multiple input images.
2. The computer of claim 1 , the instructions including further instructions to, when the first object is determined to not correspond to an anomaly, operate a vehicle based on detecting the first object by the deep neural network.
3. The computer of claim 1 , the instructions including further instructions to, when the first object is determined to correspond to an anomaly, not operate a vehicle based on detecting the first object by the deep neural network.
4. The computer of claim 3 , the instructions including further instructions to input a second image including a second object to the deep neural network and, when the second object is determined to not be an anomaly, operate the vehicle based on detecting the second object by the deep neural network.
5. The computer of claim 3 , the instructions including further instructions to input a second image including a second object to the deep neural network and, when the second object is determined to be an anomaly, stop the vehicle.
6. The computer of claim 1 , the instructions including further instructions to:
determine a plurality of feature vectors and plurality of distances from the feature vector subspace for a plurality of objects included in a plurality of images input to the deep neural network;
determine a mean and a standard deviation for the plurality of distances; and
determine that the first object corresponds to an anomaly based on the first distance being greater than three standard deviations.
7. The computer of claim 1 , wherein the K-means single value decomposition algorithm determines a sparse representation of the overcomplete dictionary by determining a greedy orthogonal matching pursuit algorithm on the overcomplete dictionary.
8. The computer of claim 1 , wherein the first feature vector corresponds to an appearance of the first object.
9. The computer of claim 8 , wherein the appearance includes one or more of size, shape, texture, color, and arrangement of object features, wherein the object features are visually distinct portions of the first object.
10. The computer of claim 1 , wherein the distance is determined in a high dimensional space wherein the high dimensional space has more than three dimensions.
11. The computer of claim 1 , wherein the deep neural network is trained using unsupervised learning.
12. A method comprising:
inputting an image including a first object to a deep neural network trained to detect objects;
outputting a first feature vector corresponding to the first object from the deep neural network;
measuring a first distance from the first feature vector to a feature vector subspace determined using a K-means single value decomposition algorithm on an overcomplete dictionary of feature vectors; and
determining that the first object is an anomaly based on the first distance being greater than a bound determined based on multiple input images.
13. The method of claim 12 , further comprising, when the first object is determined to not correspond to an anomaly, operating a vehicle based on detecting the first object by the deep neural network.
14. The method of claim 12 , further comprising, when the first object is determined to correspond to an anomaly, not operating a vehicle based on detecting the first object by the deep neural network.
15. The method of claim 14 , further comprising inputting a second image including a second object to the deep neural network and, when the second object is determined to not be an anomaly, operating the vehicle based on detecting the second object by the deep neural network.
16. The method of claim 14 , further comprising inputting a second image including a second object to the deep neural network and, when the second object is determined to be an anomaly, stopping the vehicle.
17. The method of claim 12 , further comprising:
determining a plurality of feature vectors and plurality of distances from the feature vector subspace for a plurality of objects included in a plurality of images input to the deep neural network;
determining a mean and a standard deviation for the plurality of distances; and
determining that the first object corresponds to an anomaly based on the first distance being greater than three standard deviations.
18. The method of claim 12 , wherein the K-means single value decomposition algorithm determines a sparse representation of the overcomplete dictionary by determining a greedy orthogonal matching pursuit algorithm on the overcomplete dictionary.
19. The method of claim 12 , wherein the first feature vector corresponds to an appearance of the first object.
20. The method of claim 19 , wherein the appearance includes one or more of size, shape, texture, color, and arrangement of object features, wherein the object features are visually distinct portions of the first object.