Model training method and image change analysis method based on semi-supervised confrontation learning

Abstract

The present invention proposes a model training method and image change analysis method based on semi-supervised adversarial learning, wherein the training method includes: obtaining multi-temporal images to be trained; the multi-temporal images to be trained include marked image and an unlabeled image; the annotated image and the first segmentation prediction map obtained based on the annotated image are subjected to adversarial training, the adversarial training includes a semantic segmentation network and a discriminative network; the unlabeled image is input into the discriminative network after adversarial training , to obtain the trustworthy region in the unlabeled image close to the marked image, and use the trustworthy region as the supervisory signal to conduct semi-supervised training on the semantic segmentation network after adversarial training, and obtain the image feature classification model. The method and system proposed by the invention reduce the workload of data labeling, are easy to operate, and can accurately and finely classify various ground object elements in images.

Description

technical field [0001] The invention relates to the technical fields of remote sensing image processing and analysis and computer vision, in particular to a model training method and an image change analysis method based on semi-supervised confrontation learning. Background technique [0002] The 21st century is a stage of rapid urbanization. Excessive and disorderly urban expansion will produce a series of negative effects. Effective change detection in cities can analyze changes in land use and land cover (LULC), and analyze changes in different periods. The driving force of urban expansion provides reference for urban management and planning, land use protection, and predicts the future development trend of the city. Therefore, in-depth research on urban expansion changes is essential to promote the sustainable development of cities. [0003] Remote sensing change detection is a technology that uses multi-temporal remote sensing images to analyze the characteristics and i...

AI Technical Summary

Problems solved by technology

Image semantic segmentation is a classification at the pixel level. In recent years, methods based on deep learning have shown advantages in the field of semantic segmentation, but the performance of general frameworks depends on the amount of labeled data, and they all adopt a fully supervised training method, while pixel-level Labeling is not only time-consuming, laborious, and expensive, but also manual labeling is difficult to guarantee high accuracy

[0004] In order to reduce the workload of data labeling and improve the generalization ability of the semantic segmentation model, many semi-supervised and weakly supervised methods have emerged in recent years. The workload of labeling samples in datasets is heavy and requires multiple iterations of training.

In particular, high-resolution remote sensing images have complex scenes, complex surface features and rich background interference, and the workload of pixel-level labeling is more heavy. None of the existing frameworks is suitable for training and analysis of remote sensing images.

Images