SAR image based on adversarial generation network, and visible light image mode conversion method

Abstract

The invention relates to a SAR image based on adversarial generation network, and a visible light image mode conversion method. The method comprises: firstly, extracting feature vectors of satellite images at the same position, and enabling the feature vectors to serve as the prior information of the SAR image; And inputting the priori information and the SAR image into a generator to generate a visible light image with an SAR image target. Secondly, a discriminator in the generative adversarial network is trained, and a formula LGAN (GAB, D, A and B) = Eb-B [log D (b)] + Ea-A [log (1-D (GAB (a)))] is adopted as discrimination loss; And finally, judging whether the trained adversarial generation network has a model folding error or not, namely inputting different SAR images, and enabling most of the output of the generator to be the same visible light image. And meanwhile, another generator is trained, and the feature similarity of the two images is compared by adopting generation loss; generating a loss of LGAN (GAB, GBA, A, B) = Ea-A [| | GAB (GBA (a))- A||1]. And when network training is completed, curves of the judgment loss and the generation loss tend to be stable, the judgment loss is not increased any more, and the generation loss is not reduced any more.

Description

technical field [0001] The invention belongs to the field of image translation in deep learning, and relates to a mode conversion method for SAR images and visible light images based on an adversarial generation network. Background technique [0002] Since 1978, the emergence of synthetic aperture radar (SAR) has set off a magnificent revolution in radar technology. Its incomparable all-weather, all-weather and many other advantages and the resulting wide application prospects have attracted countless eyes of the radar science community. The ensuing SAR-related research constitutes the main theme of this wave of technological revolution, and SAR systems with different wave bands, different polarizations, and even different resolutions continue to emerge. There is no doubt that this great change has affected all fields of military and civilian use. [0003] Due to the improvement of resolution, the amount of SAR data increases exponentially, and artificial information proce...

AI Technical Summary

Problems solved by technology

[0003] Due to the improvement of resolution, the amount of SAR data increases exponentially, and artificial information processing and application research (such as target recognition) face many difficulties: first, in a large area, manual interpretation is required to realize ground object detection based on SAR images , Recognition tasks, the amount of tasks far exceeds the limit of rapid judgment by humans, and the resulting subjective errors and misunderstandings are inevitable

Secondly, the special imaging mechanism of the SAR image makes the target very sensitive to the azimuth angle. A large azimuth angle difference will lead to a completely different SAR image, which will further increase the visual difference between the SAR image and the optical image, and increase the The difficulty of image interpretation and judgment; again, with the continuous improvement of SAR sensor resolution and the diversification of sensor modes, bands and polarization methods, the target information in SAR images also shows explosive growth. The point target on the medium and low resolution images has become a surface target with rich detail features and scattering features. The variety and instability have greatly increased, so the traditional information processing and application methods can no longer meet the needs of practical applications, and relevant key technologies must be tackled to speed up data processing and improve the accuracy of information extraction

Images