| CPC G06V 40/16 (2022.01) [G06F 18/214 (2023.01); G06N 3/045 (2023.01)] | 20 Claims |
|
1. A training method of a neural network for object recognition, comprising:
inputting a training image set containing an object to be recognized, which includes a set of image samples and a set of variation image samples, into the neural network to extract a student feature of each of the image samples;
dividing the image samples in the training image set into simple samples and hard samples based on feature distances of the extracted student features;
for each kind of the image sample and the variation image sample:
performing, by a respective transitive transfer adapter, a transitive transfer based on the dividing on the student feature of the image sample to obtain a transferred student feature;
calculating a distillation loss of the transferred student feature of the image sample relative to a teacher feature extracted from corresponding image sample of the other kind;
classifying, by a respective classifier, the image sample based on the student feature; and
calculating a classification loss of the image sample,
calculating a total loss related to the training image set based on the distillation losses and the classification losses calculated for image samples; and
updating parameters of the neural network according to the calculated total loss.
|