Search

CN-115578588-B - Foreground target migration method for unsupervised domain self-adaption

CN115578588BCN 115578588 BCN115578588 BCN 115578588BCN-115578588-B

Abstract

The invention discloses a foreground target migration method for self-adaption of an unsupervised domain, which comprises the steps of obtaining a marked source domain sample and an unmarked target domain sample as training samples, inputting the source domain sample and the target domain sample into a deep neural network for training to obtain a classification model, and inputting data to be classified in the target domain into the classification model to obtain a classification result. The invention also provides a foreground object migration device for unsupervised domain self-adaption. The foreground target migration method and device for the unsupervised domain self-adaption can improve classification accuracy and classification efficiency.

Inventors

  • TANG XIANGYAN
  • CHENG JIEREN
  • LIU LE

Assignees

  • 海南大学

Dates

Publication Date
20260508
Application Date
20221012

Claims (4)

  1. 1. A method for unsupervised domain adaptation of foreground object migration, the method comprising: acquiring a marked source domain sample and an unmarked target domain sample as training samples; distinguishing foreground features and background features in the source domain sample based on priori knowledge, extracting the foreground features in the source domain sample, and increasing the weight of the foreground features in the source domain sample; Distinguishing foreground features and background features in the target domain sample based on priori knowledge, and extracting the foreground features in the target domain sample; Inputting the foreground features in the source domain sample and the foreground features in the target domain sample into a deep neural network, and enabling the deep neural network to perform comparison learning training on the basis of the foreground features in the source domain sample and the foreground features in the target domain sample so as to obtain a classification model; and inputting the data to be classified in the target domain into the classification model to obtain a classification result.
  2. 2. The method of claim 1, wherein performing a comparison study based on the foreground features in the source domain sample and the foreground features in the target domain sample comprises: and comparing the foreground features in the source domain sample with the foreground features in the target domain sample by adopting a class level alignment mode.
  3. 3. A foreground object migration apparatus for unsupervised domain adaptation, the apparatus comprising: The device comprises a source domain sample, an extraction module, a background feature extraction module, a priori knowledge analysis module and a target domain analysis module, wherein the source domain sample is used for obtaining a marked source domain sample and an unmarked target domain sample as training samples; the training module is used for inputting the foreground features in the source domain sample and the foreground features in the target domain sample into a deep neural network, so that the deep neural network performs comparison learning training based on the foreground features in the source domain sample and the foreground features in the target domain sample to obtain a classification model; and the classification module is used for inputting the data to be classified in the target domain into the classification model to obtain a classification result.
  4. 4. The apparatus of claim 3, wherein the extraction module is further to: and comparing the foreground features in the source domain sample with the foreground features in the target domain sample by adopting a class level alignment mode.

Description

Foreground target migration method for unsupervised domain self-adaption Technical Field The invention relates to the technical field of deep learning, in particular to a foreground target migration method for unsupervised domain self-adaption. Background Unsupervised domain adaptation is based on a deep neural network and is implemented by training a classification model from labeled source domains to adapt to unlabeled target domains. The performance of the classification model in the target domain is poor due to the different data distribution of the source domain and the target domain. In the prior art, the performance of the classification model is improved by aligning global features of the source domain and the target domain and learning the domain invariant features. In this way, however, differences between foreground and background features in the sample are ignored and structural information in the foreground object of the sample is not taken into account, i.e. the fine class-specific structure of the sample is ignored, resulting in noise prediction near the classification model boundary. In view of this, it is necessary to provide an unsupervised domain adaptive method that improves classification accuracy and classification efficiency. Disclosure of Invention The application aims to improve classification accuracy and classification efficiency. In order to achieve the above objective, the embodiments of the present invention provide a method and an apparatus for foreground object migration for unsupervised domain adaptation. The technical scheme is as follows: In a first aspect, a method for foreground object migration for unsupervised domain adaptation, the method comprising: acquiring a marked source domain sample and an unmarked target domain sample as training samples; Inputting the source domain sample and the target domain sample into a deep neural network for training to obtain a classification model; and inputting the data to be classified in the target domain into the classification model to obtain a classification result. Further, the method further comprises: Extracting foreground features in the source domain sample and extracting foreground features in the target domain sample; and comparing and learning based on the foreground features in the source domain sample and the foreground features in the target domain sample. Further, extracting foreground features in the source domain sample and extracting foreground features in the target domain sample includes: The foreground features and the background features in the source domain sample or the target domain sample are distinguished based on a priori knowledge. Further, after extracting the foreground feature in the source domain sample, the method includes: And setting the weight of the foreground characteristic in the source domain sample. Further, based on the foreground features in the source domain sample and the foreground features in the target domain sample, performing a comparison learning includes: and comparing the foreground features in the source domain sample with the foreground features in the target domain sample by adopting a class level alignment mode. In a second aspect, a foreground object migration apparatus for unsupervised domain adaptation, the apparatus comprising: the extraction module is used for acquiring marked source domain samples and unmarked target domain samples as training samples; the training module is used for inputting the source domain sample and the target domain sample into a deep neural network for training to obtain a classification model; and the classification module is used for inputting the data to be classified in the target domain into the classification model to obtain a classification result. Further, the extraction module is further configured to: Extracting foreground features in the source domain sample and extracting foreground features in the target domain sample; and comparing and learning based on the foreground features in the source domain sample and the foreground features in the target domain sample. Further, the extraction module is further configured to: The foreground features and the background features in the source domain sample or the target domain sample are distinguished based on a priori knowledge. Further, the extraction module is further configured to: And setting the weight of the foreground characteristic in the source domain sample. Further, the extraction module is further configured to: and comparing the foreground features in the source domain sample with the foreground features in the target domain sample by adopting a class level alignment mode. The technical scheme provided by the embodiment of the invention has the beneficial effects that in the self-adaptive process of the unsupervised domain, the prior knowledge is utilized to distinguish and compare the foreground features and the background features in the source domain sample and the target dom