Unsupervised domain adaptation method and system based on target domain self-supervised learning

Abstract

The invention provides an unsupervised domain adaptation method and system based on target domain self-supervised learning. The unsupervised domain adaptation method comprises a domain invariant feature extraction step, an image feature stepped domain alignment step, a semantic segmentation step, an edge generation step, a segmentation image domain alignment step, an edge image domain alignment step and an edge consistency constraint step. According to the method, effective self-supervised learning is carried out on the target field, so that the unsupervised segmentation precision of the target field is improved and good field adaptation is realized.

Description

technical field [0001] The invention relates to the fields of computer vision and image processing, in particular to an unsupervised domain adaptation method and system based on self-supervised learning of a target domain. Background technique [0002] In recent years, with the rapid development of computer vision technology, semantic segmentation task, as an important branch of vision tasks, has been widely studied. Using a deep neural network to train a semantic segmentation task often requires manual labeling of images in the dataset as supervision, but manual labeling for each dataset is extremely costly in manpower, material and financial resources. Therefore, we generally use labeled datasets for other domain-related tasks, and use their data and labels to allow the model to learn under certain supervision, thereby alleviating the model's dependence on the labeling of the target dataset. However, since datasets from different domains often have large differences in th...

AI Technical Summary

Problems solved by technology

In addition, many existing methods tend to focus on the semantic segmentation task itself, attempting to improve the accuracy of semantic segmentation through KL divergence, clustering, weighted loss functions and other methods, but they all ignore the area that affects the semantic segmentation accuracy of the target domain the most: Segment edge regions, which are generally over-segmented or under-segmented

Images