| CPC G06V 10/764 (2022.01) | 8 Claims |
|
1. A method performed by one or more computers for learning a machine learning model classifying an input image into known classes and an unknown class by extracting a feature of the input image, each of the known classes showing that the input image falls into any one of specific categories and the unknown class showing that the input image does not fall into any of the specific categories, the method comprising:
acquiring a set of features by inputting a plurality of samples of training data into the machine learning model, each of the plurality of samples of training data being with a label;
calculating a loss function based on the set of features; and
learning the machine learning model to reduce the loss function, wherein
the loss function is configured to include:
a first loss function configured to become larger as first distances become smaller than a predetermined margin, wherein the first distances are distances in a feature space between features of a plurality of first anchor points selected out of the plurality of samples of training data with the label corresponding to the unknown class and features of one or more of samples of training data; and
a second loss function configured to become larger as second distances become larger and as third distances become smaller, wherein
the second distances are distances in the feature space between features of a plurality of second anchor points selected out of the plurality of samples of training data with the label corresponding to any one of the known classes and features of one or more of samples of training data with a same label as the plurality of second anchor points, and
the third distances are distances in the feature space between the features of the plurality of second anchor points and features of one or more of samples of training data with a different label as the plurality of second anchor points.
|