A domain adaptive image semantic segmentation method based on antagonistic learning

Abstract

The invention provides a domain adaptive image semantic segmentation method based on antagonistic learning, which comprises the following main contents: domain adaptation, network structure and outputspace adaptation. The process comprises the following steps: firstly, inputting images of a source domain and a target domain to a segmentation network to predict the source domain and the target domain to obtain segmentation output; generating the segmentation loss of the source domain by the source prediction obtained from the source output. Then, two segmented outputs are used as the input ofthe discriminator to generate anti-loss, and then the anti-loss is transmitted to the segmented network. Finally, the pixel-level semantic image segmentation is achieved by minimizing the segmentationloss and maximizing the antagonistic loss. The invention develops a multi-level confrontation learning method, which can effectively align the scene layout and the local context between the source and the target images in the self-adaptation of the segmentation space. In addition, the invention is simple, convenient and easy to operate, and can well solve the influence of adapting to the complexity of the high-dimensional features.

Description

technical field [0001] The invention relates to the field of graphics and image processing, in particular to a domain-adaptive image semantic segmentation method based on adversarial learning. Background technique [0002] Image semantic segmentation refers to classifying each pixel of the image, obtaining the content of the image and the position of the target in the image from the pixel level. At present, semantic segmentation is used in underwater object detection, geographic information system, unmanned vehicle driving, medical image analysis, robotics and other fields; through training neural network, the machine can input satellite remote sensing images to automatically identify roads, rivers, crops and buildings, etc. ;In the field of intelligent medical care, semantic segmentation can be applied to tumor image segmentation, dental caries diagnosis, etc.; semantic segmentation is also the core algorithm technology of unmanned vehicle driving. After the vehicle camera ...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the problem that the adaptation of semantic segmentation features is easily affected by the complexity of high-dimensional features and low-dimensional features cannot be well adapted, the purpose of the present invention is to provide a domain-adaptive image semantic segmentation method based on adversarial learning. First, the input source domain and the image of the target domain, passed to the segmentation network to predict the source domain and the target domain to obtain the segmentation output; the source prediction obtained from the source output generates the segmentation loss of the source domain; then the two segmentation outputs are used as the input of the discriminator to generate the confrontation loss , and then pass the adversarial loss to the segmentation network; finally, by minimizing the segmentation loss and maximizing the adversarial loss, the image is segmented to meet the required pixel-level semantics

Images