An Unsupervised Domain Adaptive Semantic Segmentation Method Based on Maximum Squares Loss

Abstract

The invention discloses an unsupervised domain adaptive semantic segmentation method based on maximum square loss, comprising: (1) pre-training a semantic segmentation model on the source domain, the semantic segmentation model is based on a ResNet network; (2) simultaneously extracting semantics Segment the features of the fourth layer in the model, add an additional classification network, and perform the same pre-training on this network branch; (3) supervised training of the semantic segmentation model on the source domain, while using the maximum squares on the target domain. The semantic segmentation model is trained unsupervised with loss; (4) In the target domain, the output of the last layer of the ResNet network is used as a pseudo-label, and the fourth layer of features is trained unsupervised; (5) After the model is trained, the image is analyzed on the target domain. Output its semantic segmentation map. By using the present invention, in the semantic segmentation effect of adapting the unsupervised domain, more training can be obtained for difficult-to-train samples and small object categories, and the quality of semantic segmentation on the final target domain can be improved.

Description

technical field [0001] The invention belongs to the field of semantic segmentation of unsupervised domain adaptation, in particular to an unsupervised domain adaptive semantic segmentation method based on maximum square loss. Background technique [0002] In the past ten years, deep learning has achieved great success in semantic segmentation tasks. With a large number of publicly available semantic segmentation models available online, semantic segmentation has received extensive attention from researchers in both industry and academia. For example, semantic segmentation technology is needed in unmanned driving systems to identify signal lights and help identify obstacles. For datasets currently used for deep network training, such as the PASCAL VOC-2012 and Cityscapes datasets, researchers have made significant progress in the performance of deep models on current datasets. However, these real datasets with pixel-wise semantic labels require a lot of manual annotation ef...

AI Technical Summary

Problems solved by technology

However, in the unsupervised domain adaptation task, there are no well-annotated class labels on the target domain to compute class frequencies, making conventional weighting factors unusable in this task

Images