Back to results
US 11,455,525 B2Granted
Method and apparatus of open set recognition and a computer readable storage medium
AssigneeFujitsu Limited
InventorsXiaoyi Yu, Jun Sun
Life of this patent
Estimating the term…
Abstract
A method and apparatus of open set recognition, and a computer-readable storage medium are disclosed. The method comprises acquiring auxiliary data and training data of known categories for open set recognition, training a neural network alternately using the auxiliary data and the training data, until convergence; extracting a feature of data to be recognized for open set recognition, using the trained neural network; and recognizing a category of data to be recognized, based on the feature of the data to be recognized.
Claims16
A method of open set recognition, comprising:
- acquiring auxiliary data and training data of known categories for open set recognition;
- training a convolutional neural network alternately using the auxiliary data and the training data, until convergence occurs;
- extracting a feature of data to be recognized for the open set recognition, using the trained convolutional neural network;
- and
- recognizing a category of the data to be recognized, based on the feature of the data to be recognized,
- wherein the training the convolutional neural network comprises:
- performing the following operations iteratively, until the convolutional neural network converges, wherein an output layer of the convolutional neural network comprises a first full-connected layer and a second full-connected layer which are parallel:
- taking a predetermined amount of data out of the auxiliary data and inputting the predetermined amount of data to the convolutional neural network, calculating a loss function of the convolutional neural network taking an output of the first full-connected layer as an output of the convolutional neural network, and updating weight values in the convolutional neural network other than weight values of the second full-connected layer based on the calculated loss function;
- and
- taking a predetermined amount of data out of the training data and inputting the predetermined amount of data to the convolutional neural network, calculating a loss function of the convolutional neural network taking an output of the second full-connected layer as an output of the convolutional neural network, and updating weight values in the convolutional neural network other than weight values of the first full-connected layer based on the calculated loss function.
7 more elements
An apparatus of open set recognition, comprising:
- a memory;
- and
- a processor coupled to the memory and configured to:
- acquire auxiliary data and training data of known categories for open set recognition;
- train a convolutional neural network alternately using the auxiliary data and the training data, until convergence occurs;
- extract a feature of data to be recognized for the open set recognition, using the trained convolutional neural network;
- and
- recognize a category of the data to be recognized, based on the feature of the data to be recognized,
- wherein the processor is further configured to perform the following operations iteratively, until the convolutional neural network converges, wherein an output layer of the convolutional neural network comprises a first full-connected layer and a second full-connected layer which are parallel:
- taking a predetermined amount of data out of the auxiliary data and inputting the predetermined amount of data to the convolutional neural network, calculating a loss function of the convolutional neural network taking an output of the first full-connected layer as an output of the convolutional neural network, and updating weight values in the convolutional neural network other than weight values of the second full-connected layer based on the calculated loss function;
- and
- taking a predetermined amount of data out of the training data and inputting the predetermined amount of data to the convolutional neural network, calculating a loss function of the convolutional neural network taking an output of the second full-connected layer as an output of the convolutional neural network, and updating weight values in the convolutional neural network other than weight values of the first full-connected layer based on the calculated loss function.
9 more elements
A non-transitory computer-readable storage medium storing therein a program that can be run by a processor to perform the operations of:
- acquiring auxiliary data and training data of known categories for open set recognition;
- training a convolutional neural network alternately using the auxiliary data and the training data, until convergence occurs;
- extracting a feature of data to be recognized for the open set recognition, using the trained convolutional neural network;
- and
- recognizing a category of the data to be recognized, based on the feature of the data to be recognized,
- wherein the processor is further configured to perform the following operations iteratively, until the convolutional neural network converges, wherein an output layer of the convolutional neural network comprises a first full-connected layer and a second full-connected layer which are parallel:
- taking a predetermined amount of data out of the auxiliary data and inputting the predetermined amount of data to the convolutional neural network, calculating a loss function of the convolutional neural network taking an output of the first full-connected layer as an output of the convolutional neural network, and updating weight values in the convolutional neural network other than weight values of the second full-connected layer based on the calculated loss function;
- and
- taking a predetermined amount of data out of the training data and inputting the predetermined amount of data to the convolutional neural network, calculating a loss function of the convolutional neural network taking an output of the second full-connected layer as an output of the convolutional neural network, and updating weight values in the convolutional neural network other than weight values of the first full-connected layer based on the calculated loss function.
6 more elements
Dependencies are inferred from the claim text.