Reducing false detections for night vision cameras
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for reducing camera false detections. One of the methods includes providing, to a neural network of an image classifier that is trained to detect objects of two or more classification types, a feature vector for a respective training image; receiving, from the neural network, an output vector that indicates, for each of the two or more classification types, a likelihood that the respective training image depicts an object of the corresponding classification type; accessing, from two or more ground truth vectors each for one of the two or more classification types, a ground truth vector for the classification type of an object depicted in the training image; and adjusting one or more weights in the neural network using the output vector and the ground truth vector; and storing, in a memory, the image classifier.
1. A computer-implemented method comprising:
for each of a plurality of training images each of which are associated with a classification type from two or more classification types, the plurality of training images including at least one image for each of the two or more classification types:
providing, to a neural network of an image classifier that is trained to detect objects of the two or more classification types, a feature vector for the respective training image;
receiving, from the neural network, an output vector that indicates, for each of the two or more classification types, a likelihood that the respective training image depicts an object of the corresponding classification type;
accessing, from two or more ground truth vectors each for one of the two or more classification types, a ground truth vector for the classification type of an object depicted in the respective training image, wherein a first ground truth vector from the two or more ground truth vectors is different than a second ground truth vector from the two or more ground truth vectors. at least one ground truth vector from the two or more ground truth vectors corresponds to two training images from the plurality of training images, and both of the two training images have the same classification type;
computing a combination of the output vector and the ground truth vector for the classification type of the objected depicted in the respective training image; and
adjusting one or more weights in the neural network using the combination of the output vector and the ground truth vector for the classification type of the objected depicted in the respective training image; and
storing, in a memory, the image classifier that includes the neural network for use by a camera to classify objects detected in one or more images captured by the camera.
2. The method of claim 1 , wherein a ground truth vector for an image represents a ground truth label for the image.
3. The method of claim 1 , wherein each ground truth vector comprises a value for each of the two or more classification types.
4. The method of claim 3 , wherein each ground truth vector comprises a highest value for the classification type from the two or more classification types of the object depicted in a corresponding image.
5. The method of claim 3 , wherein:
the output has a first dimension that is the same as a second dimension for the ground truth vector; and
adjusting the one or more weights in the neural network comprises:
combining, for each of the values in the ground truth vector, the respective value from the ground truth vector with a corresponding value from the output vector to generate combined values;
generating a training value using the combined values; and
adjusting the one or more weights in the neural network using the training value.
6. The method of claim 3 , wherein:
each of the two or more classification types is a classification for either an object of interest or not an object of interest; and
a ground truth vector from the two or more ground truth vectors comprises i) two or more non-negative values each of which are less than one and the sum of which equals one, each of the two or more non-negative values corresponding to a respective one of the two or more classification types, and ii) a negative value for a different classification type from the two or more classification types.
7. The method of claim 6 , wherein each of the two or more non-negative values in the ground truth vector are the same value.
8. The method of claim 1 , wherein storing, in a memory, the image classifier comprises:
combining the neural network with a binary classifier layer to generate a binary neural network trained to receive a feature vector for an image as input and output a value that indicates whether an object depicted in the image is an object of interest or is not an object of interest; and
storing, in a memory, the image classifier that includes the binary neural network for use by a camera to classify objects detected in one or more images captured by the camera.
9. The method of claim 1 , comprising providing the image classifier to a camera for use by the camera classifying objects detected in one or more images captured by the camera.
10. The method of claim 9 , wherein:
at least some training images from the plurality of training images comprise images captured in a low light environment; and
providing the image classifier to the camera comprises providing, to an infrared camera, an infrared image classifier that includes the neural network trained using the images captured in the low light environment.
11. The method of claim 1 , wherein accessing the ground truth vector comprises accessing, from three or more ground truth vectors each for one of three or more classification types, the ground truth vector for the classification type of an object depicted in the respective training image.
12. The method of claim 11 , wherein:
the three or more classification types comprise five classification types including a background classification, a spider web classification, a human classification, an animal classification, and a vehicle classification; and
accessing the ground truth vector comprises accessing, from five ground truth vectors each for one of the five classification types, the ground truth vector for the classification type of an object depicted in the respective training image.
13. A system comprising one or more computers and one or more storage devices on which are stored instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
for each of a plurality of training images each of which are associated with a classification type from two or more classification types, the plurality of training images including at least one image for each of the two or more classification types:
providing, to a neural network of an image classifier that is trained to detect objects of the two or more classification types, a feature vector for the respective training image;
receiving, from the neural network, an output vector that indicates, for each of the two or more classification types, a likelihood that the respective training image depicts an object of the corresponding classification type;
accessing, from two or more ground truth vectors each for one of the two or more classification types, a ground truth vector for the classification type of an object depicted in the respective training image, wherein a first ground truth vector from the two or more ground truth vectors is different than a second ground truth vector from the two or more ground truth vectors, at least one ground truth vector from the two or more ground truth vectors corresponds to two training images from the plurality of training images, and both of the two training images have the same classification type;
computing a combination of the output vector and the ground truth vector for the classification type of the objected depicted in the respective training image; and
adjusting one or more weights in the neural network using the combination of the output vector and the ground truth vector for the classification type of the objected depicted in the respective training image; and
storing, in a memory, the image classifier that includes the neural network for use by a camera to classify objects detected in one or more images captured by the camera.
14. The system of claim 13 , wherein a ground truth vector for an image represents a ground truth label for the image.
15. The system of claim 13 , wherein each ground truth vector comprises a value for each of the two or more classification types.
16. The system of claim 15 , wherein each ground truth vector comprises a highest value for the classification type from the two or more classification types of the object depicted in a corresponding image.
17. The system of claim 15 , wherein:
the output has a first dimension that is the same as a second dimension for the ground truth vector; and
adjusting the one or more weights in the neural network comprises:
combining, for each of the values in the ground truth vector, the respective value from the ground truth vector with a corresponding value from the output vector to generate combined values;
generating a training value using the combined values; and
adjusting the one or more weights in the neural network using the training value.
18. The system of claim 15 , wherein:
each of the two or more classification types is a classification for either an object of interest or not an object of interest; and
a ground truth vector from the two or more ground truth vectors comprises i) two or more non-negative values each of which are less than one and the sum of which equals one, each of the two or more non-negative values corresponding to a respective one of the two or more classification types, and ii) a negative value for a different classification type from the two or more classification types.
19. The system of claim 18 , wherein each of the two or more non-negative values in the ground truth vector are the same value.
20. The system of claim 13 , wherein storing, in a memory, the image classifier comprises:
combining the neural network with a binary classifier layer to generate a binary neural network trained to receive a feature vector for an image as input and output a value that indicates whether an object depicted in the image is an object of interest or is not an object of interest; and
storing, in a memory, the image classifier that includes the binary neural network for use by a camera to classify objects detected in one or more images captured by the camera.
21. A non-transitory computer storage medium encoded with instructions that, when executed by one or more computers, cause the one or more computers to perform operations comprising:
for each of a plurality of training images each of which are associated with a classification type from two or more classification types, the plurality of training images including at least one image for each of the two or more classification types:
providing, to a neural network of an image classifier that is trained to detect objects of the two or more classification types, a feature vector for the respective training image;
receiving, from the neural network, an output vector that indicates, for each of the two or more classification types, a likelihood that the respective training image depicts an object of the corresponding classification type;
accessing, from two or more ground truth vectors each for one of the two or more classification types, a ground truth vector for the classification type of an object depicted in the respective training image, wherein:
the two or more classification types comprise five classification types including a background classification, a spider web classification, a human classification, an animal classification, and a vehicle classification; and
accessing the ground truth vector comprises accessing, from five ground truth vectors each for one of the five classification types, the ground truth vector for the classification type of an object depicted in the respective training image; and
adjusting one or more weights in the neural network using a combination of the output vector and the ground truth vector for the classification type of the objected depicted in the respective training image; and
storing, in a memory, the image classifier that includes the neural network for use by a camera to classify objects detected in one or more images captured by the camera.