Search

CN-122024006-A - Method and system for enhancing learning and characterization of partial marks based on alignment of characterization differences

CN122024006ACN 122024006 ACN122024006 ACN 122024006ACN-122024006-A

Abstract

The invention discloses a bias mark learning characterization enhancement method and system based on characterization difference alignment. The method comprises the steps of estimating the relation between visual categories of different image samples in a label-independent mode by using a label distribution difference metric based on KL divergence, carrying out alignment regularization on the visual category characterization difference in a high-dimensional image characterization space output by a neural network image encoder, carrying out dimension reduction sampling on high-dimensional image characterization vectors by combining a random characterization dimension mask strategy to reduce the calculation complexity of image characterization difference alignment, seamlessly integrating a characterization difference alignment framework with the existing bias mark learning image classification method based on a depth neural network, and carrying out iterative updating on model parameters for image feature extraction. The method solves the problem that the differences among the categories in the characterization space are inconsistent in the existing bias mark learning method, and effectively improves the characterization quality and the classification performance of the bias mark learning model.

Inventors

  • LV JIAQI
  • LU XINGYU

Assignees

  • 东南大学

Dates

Publication Date
20260512
Application Date
20260113

Claims (9)

  1. 1. The bias mark learning characterization enhancement method based on characterization difference alignment is used for image classification and is characterized by comprising the following steps of: (1) Estimating the visual category relationship between different image samples in a tag-independent manner by using a tag distribution difference metric based on KL divergence based on a partial mark data set consisting of a plurality of image samples and a candidate tag set thereof; (2) Based on the visual inter-category relation estimated in the step (1), carrying out alignment regularization on visual inter-category characterization differences in a high-dimensional image characterization space output by a neural network image encoder; (3) Combining a random characterization dimension mask strategy, performing dimension reduction sampling on the high-dimension image characterization vector so as to reduce the computational complexity of image characterization difference alignment; (4) The method comprises the steps of seamlessly integrating a characterization difference alignment frame with an existing bias mark learning image classification method based on a deep neural network, and iteratively updating model parameters for image feature extraction so as to improve the classification accuracy of an image classification model on unknown image samples.
  2. 2. The method for enhancing the learning of the characterization of the partial mark based on the alignment of the characterization differences according to claim 1, wherein the implementation process of the step (1) is as follows: the tag distribution difference metric based on KL divergence is defined as: Wherein, the For the sample And The label distribution difference metric value between them, Represent the first Sample at training iteration Tags distributed in categories The value of the above-mentioned value, Represent the first Sample at training iteration Tags distributed in categories The value of the above-mentioned value, In order to adjust the super-parameters of the device, Is the total number of categories; By confidence threshold Filtering the low confidence samples, defining the weights of the pairs of samples in the disparity alignment penalty as: 。
  3. 3. The method for enhancing the learning characterization of the partial mark based on the alignment of the characterization differences according to claim 1, wherein the alignment regularization of the characterization differences among the visual categories adopts a relative difference form, and a regularization loss function is defined as: Wherein, the For the first order of characterizing the differences, For the second order of characterizing the difference, In order for the encoder parameters to be the same, In order to obtain the number of samples, For sample pairs And The final weight in the disparity alignment penalty, For sample pairs And Final weight in the disparity alignment penalty.
  4. 4. A bias marker learning characterization enhancement method based on characterization difference alignment according to claim 3, wherein the first order characterization difference is the euclidean distance between two sample characterizations: Wherein, the In the case of an encoder, In order for the encoder parameters to be the same, And For two input samples; the second order characterization difference is the absolute difference between the first order differences: 。
  5. 5. The bias mark learning characterization enhancement method based on characterization difference alignment according to claim 1, wherein the implementation procedure of the step (3) is as follows: Random characterization dimension mask strategy pass mask vector Implementation in which , For the original characterization dimension, To reduce the post-dimension, the masked first order token differences are redefined as: Wherein, the Representing element-by-element multiplication.
  6. 6. The method for enhancing the learning of the characterization by the bias marks based on the alignment of the characterization differences according to claim 1, wherein the implementation process of the step (4) is as follows: Integrating the characterization of all the supervision views into a unified characterization space, and performing inter-category difference alignment in the space; The final total loss function learns the weighted sum of the original loss and the characteristic difference alignment loss for the bias mark: Wherein, the For the original loss function of the selected bias marker learning method, In order to balance the super-parameters of the two losses, Is a regularized loss function.
  7. 7. A bias marker learning characterization enhancement system based on characterization difference alignment employing the method of any of claims 1 to 6, comprising: The relation estimation module is used for inputting a partial mark data set formed by a plurality of image samples and a candidate label set thereof, and estimating the relation between visual categories of different image samples in a label-independent mode by using a label distribution difference metric based on KL (moment-dependent interference) divergence; the difference alignment module is used for carrying out alignment regularization on the representation differences among the visual categories in the high-dimensional image representation space output by the neural network image encoder based on the relation among the visual categories; The dimension mask module is used for carrying out dimension reduction sampling on the high-dimension image characterization vector by combining with a random characterization dimension mask strategy so as to reduce the calculation complexity of image characterization difference alignment; And the model integration module is used for seamlessly integrating the characterization difference alignment frame with the existing partial mark learning image classification method based on the deep neural network and carrying out iterative updating on model parameters for extracting image features.
  8. 8. A storage medium having stored thereon a computer program which, when executed by at least one processor, implements the steps of the bias marker learning token enhancement method based on token difference alignment of any one of claims 1 to 6.
  9. 9. An electronic device comprising a memory and a processor, wherein: a memory for storing a computer program capable of running on the processor; a processor for performing the steps of the bias marker learning token enhancement method based on token difference alignment as claimed in any one of claims 1 to 6 when running said computer program.

Description

Method and system for enhancing learning and characterization of partial marks based on alignment of characterization differences Technical Field The invention belongs to the field of artificial intelligence machine learning, and particularly relates to a bias mark learning characterization enhancement method and system based on characterization difference alignment. Background Biased-label learning is an important paradigm in weakly supervised learning, where each training instance is labeled with a set of candidate labels that contains only one correct label, but specifically which is unknown. The inherent ambiguity of labels presents a fundamental challenge to learning reliable decision boundaries and efficient generalization. The existing bias mark learning method is mainly divided into two types, namely a candidate label is equally treated to construct a supervision signal based on an average method, but the supervision signal is affected by serious label noise because correct and wrong candidates cannot be distinguished, and a real label of each instance is explicitly identified during training based on an identification method, so that the most advanced method in recent years mainly belongs to the types. Although these approaches have been successful in controlled benchmarking, they tend to perform poorly in more realistic scenarios (e.g., example related candidate sets). One key observation of recent progress is that high quality characterization learning plays a key role in solving tag ambiguity. If semantically similar instances are clustered together in token space and different classes are well separated, then the model is more likely to infer the correct label based on the neighborhood structure. This insight inspires a series of token driven biased marker learning methods. However, these approaches focus primarily on local objectives (e.g., minimizing intra-class variance or maximizing inter-class separation), without explicitly modeling global structures that characterize inter-class relationships in space. Such local optimization strategies, while beneficial, tend to result in unbalanced class separation in the characterization space, with the possibility of significantly different degrees of separation between different class pairs, some being over-pushed and others remaining relatively close. This structural imbalance may compromise semantic consistency of the learning token, thereby affecting generalization performance. Disclosure of Invention Aiming at the problems that the differences among the categories of the neural network models such as ResNet, convNet in the characteristic space are inconsistent and the classification generalization performance is influenced in the image classification task of the existing partial mark learning method, the invention provides a partial mark learning characteristic enhancement method and a partial mark learning characteristic enhancement system based on characteristic difference alignment, so that the partial mark learning model learns a globally balanced semantic topological structure, and the characteristic quality and the classification performance of the model are improved. The invention discloses a bias mark learning characterization enhancement method based on characterization difference alignment, which comprises the following steps: (1) Estimating the visual category relationship between different image samples in a tag-independent manner by using a tag distribution difference metric based on KL divergence based on a partial mark data set consisting of a plurality of image samples and a candidate tag set thereof; (2) Based on the visual inter-category relation estimated in the step (1), carrying out alignment regularization on visual inter-category characterization differences in a high-dimensional image characterization space output by a neural network image encoder; (3) Combining a random characterization dimension mask strategy, performing dimension reduction sampling on the high-dimension image characterization vector so as to reduce the computational complexity of image characterization difference alignment; (4) The method comprises the steps of seamlessly integrating a characterization difference alignment frame with an existing bias mark learning image classification method based on a deep neural network, and iteratively updating model parameters for image feature extraction so as to improve the classification accuracy of an image classification model on unknown image samples. Further, the implementation process of the step (1) is as follows: the tag distribution difference metric based on KL divergence is defined as: Wherein, the For the sampleAndThe label distribution difference metric value between them,Represent the firstSample at training iterationTags distributed in categoriesThe value of the above-mentioned value,Represent the firstSample at training iterationTags distributed in categoriesThe value of the above-mentioned value