CN-114067151-B - Global feature and intrinsic feature correlation-based migration learning method

CN114067151BCN 114067151 BCN114067151 BCN 114067151BCN-114067151-B

Abstract

A migration learning method based on global features and internal features includes extracting depth features of samples through a depth neural network, selecting a certain proportion of samples from a target domain as labeled target domain samples, enabling the rest samples to belong to unlabeled target domain samples, learning a base model of the target domain by means of a semi-supervised anti-migration learning method, carrying out unsupervised clustering on all samples of a source domain and unlabeled samples of the target domain according to the depth features, respectively, enabling each cluster to be called a component, matching the components of the source domain and the target domain, enabling each match to be called a pair, and fine-tuning each pair based on the base model to enable each pair to be more consistent with target domain components in the pair. Meanwhile, the integral distribution of the features and the internal correlation among the features are considered, and knowledge migration can be better realized, so that the problems of time and labor waste and the like of manual labeling in the fields of object identification and the like are better solved, and the migration efficiency is improved.

Inventors

Yi Changan
Chen haotian
Tan haishu

Assignees

佛山科学技术学院

Dates

Publication Date: 20260512
Application Date: 20211119

Claims (8)

1. A migration learning method based on global feature and intrinsic feature association is characterized in that: the method comprises the following steps: Step A, preparing a sample, dividing the picture data with the label of the current scene into a source domain, and dividing the unlabeled picture data of the new scene into a target domain; step B, extracting depth features, namely extracting the depth features of the sample through a depth neural network; step C, selecting samples, namely randomly selecting a preset number of samples from a target domain, and labeling the selected samples; In said step C, including marking said source domain as Marking the target domain as ; Training a base model for the target domain based on the source domain in the step A, the labeled sample selected in the step C and the unlabeled sample not selected in the step C; Extracting a component, and performing unsupervised clustering on the source domain and the unselected samples without labels in the step C, wherein each cluster is the component; step F, component matching, namely matching the components in the target domain with corresponding source domain components, wherein each component in the target domain and the corresponding source domain components form a pair; In step F, including component matching using bipartite graph, measuring a distance between two components using MMD distance, including: Source domain And a target domain Respectively forming two subsets in the bipartite graph, wherein the two subsets are mutually disjoint; Acquisition of target fields Any of the components and source domain MMD distances of all components are used as weights of corresponding edges in the bipartite graph; for the target domain Selecting the least weighted edge to match the corresponding source domain Is a component of (a); And G, fine-tuning the base model, and fine-tuning the base model based on each pairing.
2. The method for migration learning based on global feature and intrinsic feature association according to claim 1, wherein: the target domain Divided into , wherein, Representing a randomly selected sample, the label of the randomly selected sample is known, A label representing the unselected samples is unknown; the said The number of samples is much smaller than Is a sample number of (a) in a sample.
3. The method for migration learning based on global feature and intrinsic feature association according to claim 2, wherein: in said step D, including based on the tagged source domain Randomly selected samples And unselected samples To be the target domain Training a base model; The training base model comprises two overlapping steps of countermeasure generation and countermeasure training, wherein based on the countermeasure generation step, each original sample generates a linking sample, and the linking sample is used for filling a source domain And a target domain And training the classifier C and the discriminator D based on the countermeasure training step.
4. A method of transition learning based on global and intrinsic feature correlations as claimed in claim 3, wherein: executing the countermeasure generation step according to the first formula and the second formula; -equation one; -formula two; Wherein: A concatenation sample representing a source domain; a splice sample representing a target domain; w represents the number of iterations of the adaptive samples, and when w=0, Representing the original samples corresponding to the source domain, Representing an original sample corresponding to the target domain; And Respectively representing balance parameters of the formula I and the formula II; model parameters representing classifier C; model parameters representing the discriminators; representing the loss function of classifier C; Representing the loss function of the arbiter D.
5. The method for transition learning based on global feature and intrinsic feature association according to claim 4, wherein: In the countermeasure generating step, constraining by using Euclidean distance (Euclidean distance) to enable the characteristic distribution of the link sample to be between the source domain and the target domain; wherein, in the formula one, Representing euclidean distance-based constraints between the original samples and the concatenated samples of the source domain; in the second formula of the present invention, Representing euclidean distance-based constraints between the original samples and the concatenated samples of the target domain.
6. A method of transition learning based on global and intrinsic feature correlations as claimed in claim 3, wherein: Executing the countermeasure training step according to the formula III and the formula IV; -formula three; -formula four; Wherein: y represents a real label; x represents the original sample; Representing a splice sample; C represents a classifier; Representing a source domain original sample with a label; Representing a labeled target domain original sample; a concatenation sample representing a source domain and a target domain; Representing cross entropy; model parameters representing classifier C; representing the loss function of classifier C; Representing the expected value of the function; Representing countermeasure training; refers to the classification loss function associated with the joined samples.
7. The method for migration learning based on global feature and intrinsic feature association according to claim 6, wherein: further comprising obtaining labeled raw samples using classifier C and arbiter D And Is a loss of (2); wherein the loss of the classifier C is formed by the original sample x and the adaptive sample Composition; Involving obtaining unlabeled raw samples using a discriminator D The loss calculation formula is shown as a formula five and a formula six; -formula five; -formula six; Wherein: Representing the loss function of the arbiter D; Model parameters representing the arbiter D; D represents a discriminator D; Representing all original samples containing source and target fields; Representing all original samples comprising the source domain and the target domain as corresponding joined samples; representing a function desire; representing a single original sample belonging to a source domain; representing a single original sample belonging to the target domain; Representing a single splice sample belonging to a source domain; representing a single splice sample belonging to a target domain; Representing countermeasure training; Representing a discriminant loss function associated with the concatenated samples.
8. The method for migration learning based on global feature and intrinsic feature association according to claim 2, wherein: in said step E, the extraction component includes identifying labeled source fields based on depth characteristics of the sample And unselected samples Performing unsupervised clustering; after clustering, samples within the same component contain at least one tag.

Description

Global feature and intrinsic feature correlation-based migration learning method Technical Field The invention relates to the technical field of artificial intelligence transfer learning, in particular to a transfer learning method based on global feature and intrinsic feature association. Background Deep learning has been very successful in the field of image classification, but its success is premised on the model needing to learn a large number of parameters and supporting a large amount of tagged data. However, once a new scene is reached (or the task is slightly changed), the original learned model often cannot be directly applied, and at this time, it is often necessary to train a new model for the new scene. In the practical classification problem, it is difficult to obtain large-scale data required for constructing a model, and the efficiency of manually labeling the data is very low. The model trained by using a large amount of manually marked data is difficult to obtain ideal effects in real data. Thus, migration learning is utilized to solve such problems. The existing image classification migration methods are all based on global class information for migration, and the similarity and difference of the intrinsic fine features of the sample are not considered, so that the performance of the model is greatly reduced, and the problem of low migration efficiency is caused. Disclosure of Invention Aiming at the defects in the background technology, the invention provides a migration learning method based on the correlation of global features and intrinsic features, and the technical scheme is applied to the field of image classification learning, and firstly, the depth features of a sample are extracted through a depth neural network; selecting a certain proportion of samples from a target domain, taking the samples as labeled target domain samples, taking the rest samples as unlabeled target domain samples, adopting a semi-supervised anti-migration learning method to learn a base model of the target domain, carrying out unsupervised clustering on all samples of a source domain and unlabeled samples of the target domain according to depth characteristics, thereby gathering samples with similar internal characteristics together, each cluster is called a component, matching the components of the source domain and the target domain, each match is called a pair, and carrying out fine tuning on each pair based on the base model to enable the samples to be more consistent with target domain components in the pair. Meanwhile, the integral distribution of the features and the internal correlation among the features are considered, and knowledge migration can be better realized, so that the problems of time and labor waste and the like of manual labeling in the fields of object identification and the like are better solved, and the migration efficiency is improved. In order to solve the above problems, the present invention provides a method for migration learning based on global feature and intrinsic feature association, comprising the following steps: Step A, preparing a sample, dividing the picture data with the label of the current scene into a source domain, and dividing the unlabeled picture data of the new scene into a target domain; step B, extracting depth features, namely extracting the depth features of the sample through a depth neural network; step C, sample selection, namely randomly selecting a preset number of samples from a target domain, and acquiring labels of the selected samples; training a base model for the target domain based on the source domain in the step A, the labeled sample selected in the step C and the unlabeled sample not selected in the step C; Extracting a component, and performing unsupervised clustering on the source domain and the unselected samples without labels in the step C, wherein each cluster is the component; step F, component matching, namely matching the components in the target domain with corresponding source domain components, wherein each component in the target domain and the corresponding source domain components form a pair; And G, fine-tuning the base model, and fine-tuning the base model based on each pairing. Preferably, in the step C, the method includes marking the source domain as D S and the target domain as D T; Dividing the target domain D T into D T＝{DTL∪DTU, wherein D TL represents randomly selected samples, the labels of the randomly selected samples are known, and D TU represents unselected samples, the labels of the unselected samples are unknown; the number of samples of D TL is much smaller than the number of samples of D TU. Preferably, in said step D, training a base model for the target domain D T based on the labeled source domain D S, the randomly selected samples D TL and the unselected samples D TU; The training base model includes two overlapping steps of countermeasure generation and countermeasure training, wherein one joint samp