Double-domain adaptive module pyramid network and unsupervised domain adaptive image segmentation method

Abstract

The invention discloses a dual-domain adaptive modular pyramid network. The network comprises an encoder, a pyramid type pooling module, a decoder, an image level domain adaptation module and a feature level domain adaptation module, the pyramid type pooling module is connected behind the encoder, the decoder is connected behind the pyramid type pooling module, and the encoder is connected with aconvolution layer with the same size as the decoder through a skip connection technology. According to the method, through adversarial training of images and feature levels, a target domain label is not needed during training, the test or use process works like a normal segmentation network, and an image level domain adaptation module and a feature level domain adaptation module are not needed; the method can improve the segmentation precision of the segmentation neural network in an unmarked new image, provides technical support for the recognition of streetscape and computer-aided diagnosisof robots and vehicles, is high in practicality, and is higher in popularization and application values.

Description

technical field [0001] The invention relates to the technical fields of digital image processing and computer vision, in particular to an unsupervised domain adaptive image segmentation method based on a dual domain adaptive module pyramid network. Background technique [0002] Digital image segmentation technology has important applications in robotics, automatic navigation, medical imaging and other fields. The traditional semantic image segmentation model based on deep learning needs to use labeled training data for supervised learning, and it is necessary to ensure that the training and test data are roughly similar, that is, have the same distribution (or be in the same domain), however, in practice It is difficult to ensure that the training data and test data or new data (actually running data) are in the same domain, which makes the model experience significant performance degradation on the test data. The traditional way to solve such problems is to label new data ...

AI Technical Summary

Problems solved by technology

However, obtaining accurate labels is cumbersome and time-consuming, and it is impossible to guarantee sufficient labels for new data. Therefore, it is necessary to design an unsupervised image segmentation technology based on a dual-domain adaptive module pyramid network. In the case of , the difference between the two domains of the new unlabeled data and the original training labeled data is reduced through the image-level and feature-level domain adaptation modules to improve the performance of the model on the new data

Images