Method, electronic device, and computer program product for dataset updating

1 . A method for dataset updating, comprising:

acquiring a reference dataset comprising a plurality of reference images, wherein each reference image has a soft label, and the soft label indicates a probability that the each reference image belongs to each target class in a plurality of target classes;

determining, according to a similarity between an input image received and the each reference image, an image soft label for the input image, determining a number of original images corresponding to a distilled reference image corresponding to each probability value and determining, based on the each probability value and the number of the original images corresponding to the each probability value, a probability that the each distilled reference image is under a corresponding target class; and

adding the input image and the image soft label for the input image to the reference dataset to update the reference dataset.

2 . The method according to claim 1 , wherein determining, according to a similarity between an input image received and the each reference image, an image soft label for the input image comprises:

determining a weight for the each reference image according to the similarity between the input image and the each reference image; and

determining the image soft label for the input image according to the soft label for the each reference image and the weight.

3 . The method according to claim 2 , wherein determining the image soft label for the input image according to the soft label for the each reference image and the weight comprises:

weighting the soft label for the each reference image using the weight to obtain a weighted soft label for the each reference image; and

summing probabilities for corresponding target classes among a plurality of the weighted soft labels to determine the image soft label for the input image.

4 . The method according to claim 1 , wherein the reference dataset comprises a distilled dataset, the distilled dataset comprising a plurality of distilled reference images obtained after a distillation operation on an original dataset, with a soft label for each distilled reference image being obtained based on a probability that each original image in the original dataset belongs to each target class in the plurality of target classes.

5 . The method according to claim 4 , further comprising:

acquiring the original dataset, wherein the original dataset comprises a plurality of original images;

performing the distillation processing on the original dataset using a capsule neural network model to acquire the distilled dataset,

wherein the soft label indicates a probability that the corresponding distilled reference image belongs to the each target class in the plurality of target classes.

6 . The method according to claim 5 , wherein the input image further comprises a corresponding image hard label, the method further comprising:

fine-tuning the capsule neural network model based on the image soft label and/or the image hard label for the input image.

7 . The method according to claim 5 , wherein performing the distillation processing on the original dataset to acquire the distilled dataset comprises:

inputting each original image in the original dataset to the capsule neural network model and acquiring a set of output features;

acquiring a set of capsule features of the capsule neural network model; and

acquiring the distilled dataset based on the set of output features and the set of capsule features.

8 . The method according to claim 7 , wherein the set of output features comprises an output feature matrix, and wherein an element in a column vector of the output feature matrix indicates a probability that each original image in the original dataset belongs to a corresponding target class in the plurality of target classes.

9 . The method according to claim 8 , wherein performing the distillation processing on the original dataset to acquire the distilled dataset comprises:

acquiring a column vector in the output feature matrix that corresponds to the corresponding target class;

determining a probability value corresponding to each distilled reference image based on elements in the column vector;

determining, based on the probability value corresponding to the each distilled reference image, a probability that the each distilled reference image is under the corresponding target class as a soft label for the distilled reference image.

10 . The method according to claim 5 , further comprising:

training the capsule neural network model based on the original dataset,

wherein the loss function for training the capsule neural network model is related to a difference of the following parameters:

mutual information between the original dataset and the distilled dataset; and

a similarity between the original dataset and the distilled dataset.

11 . An electronic device for dataset updating, the electronic device comprising:

at least one processor; and

a memory coupled to the at least one processor and having instructions stored thereon, wherein the instructions, when executed by the at least one processor, cause the electronic device to perform actions comprising:

acquiring a reference dataset comprising a plurality of reference images, wherein each reference image has a soft label, and the soft label indicates a probability that the each reference image belongs to each target class in a plurality of target classes;

determining, according to a similarity between an input image received and the each reference image, an image soft label for the input image, determining a number of original images corresponding to a distilled reference image corresponding to each probability value and determining, based on the each probability value and the number of the original images corresponding to the each probability value, a probability that the each distilled reference image is under a corresponding target class; and

adding the input image and the image soft label for the input image to the reference dataset to update the reference dataset.

12 . The electronic device according to claim 11 , wherein determining, according to a similarity between an input image received and the each reference image, an image soft label for the input image comprises:

determining a weight for the each reference image according to the similarity between the input image and the each reference image; and

determining the image soft label for the input image according to the soft label for the each reference image and the weight.

13 . The electronic device according to claim 12 , wherein determining the image soft label for the input image according to the soft label for the each reference image and the weight comprises:

weighting the soft label for the each reference image using the weight to obtain a weighted soft label for the each reference image; and

summing probabilities for corresponding target classes among a plurality of the weighted soft labels to determine the image soft label for the input image.

14 . The electronic device according to claim 11 , wherein the reference dataset comprises a distilled dataset, the distilled dataset comprising a plurality of distilled reference images obtained after a distillation operation on an original dataset, with a soft label for each distilled reference image being obtained based on a probability that each original image in the original dataset belongs to each target class in the plurality of target classes.

15 . The electronic device according to claim 14 , wherein the instructions, when executed by the at least one processor, cause the electronic device to perform the following:

acquiring the original dataset, wherein the original dataset comprises a plurality of original images;

performing the distillation processing on the original dataset using a capsule neural network model to acquire the distilled dataset,

wherein the soft label indicates a probability that the corresponding distilled reference image belongs to the each target class in the plurality of target classes.

16 . The electronic device according to claim 15 , wherein the input image further comprises a corresponding image hard label, and the instructions, when executed by the at least one processor, cause the electronic device to perform the following:

fine-tuning the capsule neural network model based on the image soft label and/or the image hard label for the input image.

17 . The electronic device according to claim 15 , wherein performing the distillation processing on the original dataset to acquire the distilled dataset comprises:

inputting each original image in the original dataset to the capsule neural network model and acquiring a set of output features;

acquiring a set of capsule features of the capsule neural network model; and

acquiring the distilled dataset based on the set of output features and the set of capsule features.

18 . The electronic device according to claim 17 , wherein the set of output features comprises an output feature matrix, and wherein an element in a column vector of the output feature matrix indicates a probability that each original image in the original dataset belongs to a corresponding target class in the plurality of target classes.

19 . A non-transitory machine-readable storage medium containing instructions, wherein the instructions, when executed, cause a processor to perform operations to:

acquire a reference dataset comprising a plurality of reference images, wherein each reference image has a soft label, and the soft label indicates a probability that the each reference image belongs to each target class in a plurality of target classes;

determine, according to a similarity between an input image received and the each reference image, an image soft label for the input image, determining a number of original images corresponding to a distilled reference image corresponding to each probability value and determining, based on the each probability value and the number of the original images corresponding to the each probability value, a probability that the each distilled reference image is under a corresponding target class; and

add the input image and the image soft label for the input image to the reference dataset to update the reference dataset.