Generative adversarial migration learning method based on sketch annotation information

Abstract

The invention discloses a generative adversarial migration learning method based on sketch annotation information. The generative adversarial migration learning method comprises the following steps: obtaining an initial sketch map, and constructing a paired data set with a source domain image-source domain image edge annotation map form; constructing an edge segmentation depth network based on sketch annotation information and training; selecting a target domain sample based on the matrix norm; constructing and training a generative adversarial migration learning network based on sketch annotation information, wherein the network comprises a depth generator network, a depth discriminator network, an edge segmentation depth network based on sketch annotation information and a depth classifier network; inputting a target domain image to obtain a classification result of the target domain image; and utilizing the similarity of the source domain data structure and the target domain data structure, and through structural constraint, generating the samples which determine the labels and accord with target domain distribution, so as to carry out label transmission, and realize cross-domain classification. The classification accuracy is improved, and the cross-domain classification task is realized.

Description

technical field [0001] The invention belongs to the technical field of image classification, and in particular relates to a generation confrontation transfer learning method based on sketch annotation information, which can be used for cross-domain image classification. Background technique [0002] Deep learning has achieved remarkable results in image classification. Under the traditional deep learning framework, the learning task is to learn a classification network on a given labeled training data set. The more parameters the model has, the more complex it is. However, with the advent of the era of big data, the cost of obtaining data is getting lower and lower, but the cost of calibrating data labels has not been reduced, making the deep learning network in processing this A blockage was encountered with some data. Deep transfer learning breaks the traditional framework, uses the data domains that have been marked, and transfers knowledge by looking for similarities be...

AI Technical Summary

Problems solved by technology

The problem with this type of method is that due to the abstraction of the feature space, it is actually impossible to judge whether the extracted features are domain-invariant features.

Images