Label generation method, image classification method, and electronic device
The present application discloses a label generation method including identifying an unlabeled image by a teacher model and a student model respectively to obtain first classification reference information of the teacher model and second classification reference information of the student model. A first number of detection objects of each category in the teacher model is determined according to the first classification reference information, and a second number of detection objects of each category in student model is determined according to the second classification reference information. Once a probability threshold of each category is determined according to the first number of detection objects and the second number of detection objects, a pseudo label of the unlabeled image is generated for the student model according to the first classification reference information and a probability threshold of each predicted category.
1 . A label generation method comprising:
obtaining second classification reference information of a teacher model via identifying an unlabeled image by the teacher model, and obtaining first classification reference information of a student model via identifying the unlabeled image by the student model; the first classification reference information and the second classification reference information comprising detection objects of the unlabeled image and a probability of each of the detection objects in a predicted category;
determining a number of detection objects of each predicted category in the teacher model according to the second classification reference information, and determining a number of detection objects of each predicted category in the student model according to the first classification reference information;
determining a probability threshold of each predicted category according to the number of detection objects of each predicted category in the teacher model and the number of detection objects of each predicted category in the student model;
generating a pseudo label of the unlabeled image for the student model according to the second classification reference information and the probability threshold of each predicted category.
2 . The method according to claim 1 , wherein determining the probability threshold of each predicted category according to the number of detection objects of each predicted category in the teacher model and the number of detection objects of each predicted category in the student model comprises:
determining a learning rate of each predicted category according to the number of detection objects of each predicted category in the teacher model and the number of detection objects of each predicted category in the student model;
determining the probability threshold of each predicted category according to the learning rate of each predicted category.
3 . The method according to claim 2 , wherein determining the learning rate of each predicted category according to the number of detection objects of each predicted category in the teacher model and the number of detection objects of each predicted category in the student model comprises:
determining a first proportion of detection objects of each predicted category according to a ratio between the number of detection objects of each predicted category in the teacher model and a first total number of the detection objects, the first total number of detection objects representing a sum of the number of detection objects of each predicated category in the teacher model;
determining a second proportion of detection objects of each predicted category according to a ratio between the number of detection objects of each predicted category in the student model and a second total number of detection objects, the second total number of detection objects representing a sum of the number of detection objects of each predicated category in the student model;
determining the learning rate of each predicted category according to a ratio between the first proportion of detection objects and the second proportion of detection objects.
4 . The method of claim 2 , wherein determining the probability threshold of each predicted category according to the learning rate of each predicted category comprises:
determining a maximum learning rate from the learning rates;
determining a weight coefficient of each predicted category according to a ratio between the learning rate and the maximum learning rate;
determining the probability threshold of each predicted category according to a product of the weight coefficient and a preset probability upper limit value.
5 . The method according to claim 1 , wherein determining the number of detection objects of each predicted category in the teacher model according to the second classification reference information, and determining the number of detection objects of each predicted category in the student model according to the first classification reference information comprises:
determining the predicted category of a maximum probability as a predicted category of each detection object corresponding to the teacher model, and determining the number of detection objects of each predicted category in the teacher model according to the predicted category of each detection object corresponding to the teacher model;
determining the predicted category of the maximum probability as the predicted category of each detection object corresponding to the student model, and determining the number of detection objects of each predicted category in the student model according to the predicted category of each detection object corresponding to the student model.
6 . The method according to claim 5 , comprising:
in response that the predicted category of each detection object corresponding to the teacher model is greater than the probability threshold of the predicted category, generating the pseudo label of the unlabeled image for the student model according to the predicted category of each detection object corresponding to the teacher model.
7 . An image classification method, comprising:
identifying an image to be processed by a target model of an image classification model, and obtaining classification reference information of the target model;
determining a detection object comprised in the image to be processed and a category of the detection object according to the classification reference information of the target model;
wherein the image classification model is a model with a semi-supervised object detection framework, and the image classification model comprises a teacher model and a student model, the target model comprises the teacher model and/or the student model;
the image classification model is obtained by training according to target classification reference information of a labeled image in an image collection, target classification reference information of an unlabeled image in the image collection, a category label of the labeled image and a pseudo label of the unlabeled image; the target classification reference information of the labeled image is obtained via identifying the labeled image by the student model of the image classification model, and the target classification information of the unlabeled image is obtained via identifying the unlabeled image by the student model; the pseudo label of the unlabeled image is determined according to the method described in claim 1 .
8 . An electronic device, comprising:
a processor; and
a storage device storing instructions, which when executed by the processor, cause the processor to;
obtain second classification reference information of a teacher model via identifying an unlabeled image by the teacher model, and obtain first classification reference information of a student model via identifying the unlabeled image by the student model; the first classification reference information and the second classification reference information comprising detection objects of the unlabeled image and a probability of each of the detection objects in a predicted category;
determine a number of detection objects of each predicted category in the teacher model according to the second classification reference information, and determine a number of detection objects of each predicted category in the student model according to the first classification reference information;
determine a probability threshold of each predicted category according to the number of detection objects of each predicted category in the teacher model and the number of detection objects of each predicted category in the student model; and
generate a pseudo label of the unlabeled image for the student model according to the second classification reference information and the probability threshold of each predicted category.
9 . The electronic device according to claim 8 , wherein the processor determines the probability threshold of each predicted category according to the number of detection objects of each predicted category in the teacher model and the number of detection objects of each predicted category in the student model by:
determining a learning rate of each predicted category according to the number of detection objects of each predicted category in the teacher model and the number of detection objects of each predicted category in the student model; and
determining the probability threshold of each predicted category according to the learning rate of each predicted category.
10 . The electronic device according to claim 9 , wherein the processor determines the learning rate of each predicted category according to the number of detection objects of each predicted category in the teacher model and the number of detection objects of each predicted category in the student model by:
determining a first proportion of detection objects of each predicted category according to a ratio between the number of detection objects of each predicted category in the teacher model and a first total number of the detection objects, the first total number of detection objects representing a sum of the number of detection objects of each predicated category in the teacher model;
determining a second proportion of detection objects of each predicted category according to a ratio between the number of detection objects of each predicted category in the student model and a second total number of detection objects, the second total number of detection objects representing a sum of the number of detection objects of each predicated category in the student model; and
determining the learning rate of each predicted category according to a ratio between the first proportion of detection objects and the second proportion of detection objects.
11 . The electronic device of claim 9 , wherein the processor determines the probability threshold of each predicted category according to the learning rate of each predicted category by:
determining a maximum learning rate from the learning rates;
determining a weight coefficient of each predicted category according to a ratio between the learning rate and the maximum learning rate;
determining the probability threshold of each predicted category according to a product of the weight coefficient and a preset probability upper limit value.
12 . The electronic device according to claim 8 , wherein the processor determines the number of detection objects of each predicted category in the teacher model according to the second classification reference information, and determine the number of detection objects of each predicted category in the student model according to the first classification reference information by:
determining the predicted category of a maximum probability as a predicted category of each detection object corresponding to the teacher model, and determining the number of detection objects of each predicted category in the teacher model according to the predicted category of each detection object corresponding to the teacher model;
determining the predicted category of the maximum probability as a predicted category of each detection object corresponding to the student model, and determining the number of detection objects of each predicted category in the student model according to the predicted category of each detection object corresponding to the student model.
13 . The electronic device according to claim 12 , wherein the processor is further caused to:
in response that the predicted category of each detection object corresponding to the teacher model is greater than the probability threshold of the predicted category, generate the pseudo label of the unlabeled image for the student model according to the predicted category of each detection object corresponding to the teacher model.