High-resolution SAR ship image generation method based on generative adversarial network

Abstract

The invention relates to a high-resolution SAR image generation method based on a generative adversarial network. The method comprises the steps of firstly, establishing a multi-scale SAR ship image data set, sampling slices with the real image size of 256*256 to different resolutions through a down-sampling layer, and establishing a network structure of the generative adversarial network; takingthe training set of the SAR ship images as input, iterating the training model for multiple times, optimizing a target function, and testing the SAR ship images on the test set to obtain a convergentgenerative adversarial network; and finally, taking the noise and the ship category vector as input of a generator, and inputting into the convergent generative adversarial network to obtain a corresponding high-resolution SAR ship image. According to the method, a local response normalization layer is added in the generator, a network training process is stabilized, a multi-scale loss item is added, images generated in different scales are input into a discriminator, losses in different scales are calculated and finally added globally, and the quality of SAR images generated in different scales is improved.

Description

technical field [0001] The invention belongs to the technical field of high-resolution SAR image generation, and in particular relates to a high-resolution SAR ship image generation method based on a generative confrontation network. Background technique [0002] High-resolution SAR images have rich texture feature information and are not affected by clouds and rain. They are widely used in the field of remote sensing target recognition, especially in marine target recognition under extreme conditions of cloud and rain, such as ship recognition. Recently, deep learning object detection algorithms have been widely used, completely affecting the field of SAR image processing, such as the field of SAR ship recognition, but this is mainly due to the large-scale annotated training data sets of different scales. However, acquiring a large number of manually annotated SAR images of different scales is an expensive and time-consuming task. SAR image simulation is the main means to ...

AI Technical Summary

Problems solved by technology

However, acquiring a large number of manually annotated SAR images at different scales is an expensive and time-consuming task

SAR image simulation is the main means to solve the shortage of high-resolution SAR data, while traditional methods based on physical simulation models, such as Kirchhoff physical optics method, geometrical optics approximation, integral equation method or Phong model, etc. Huge memory, high cost

In this case, the image generation algorithm based on generative confrontation network (GAN) can generate realistic and multi-modal sample images, and has the advantages of low loss and end-to-end, but there are still the following shortcomings: only low-resolution SAR can be generated The goal is not to generate SAR images of multiple scales; the gradient update is unstable during the generation process, and training collapse often occurs; the generated SAR samples are only qualitatively evaluated visually, and the potential of generating SAR samples cannot be evaluated in other scenarios

[0003] When the original assisted generative confrontation network (ACGAN) generates high-resolution images, training collapses due to unstable network gradient updates

When the discriminator is overtrained, the generator tends to have gradient disappearance, which makes it difficult to reduce the loss of the generator.

And when the discriminator is undertrained, the generator will experience unstable gradient updates.

Therefore, the original ACGAN is difficult to train, and often needs to constantly adjust the parameters to find the optimal solution

At the same time, in the original ACGAN generation process, no constraints were added to the images of different scales generated in the middle, making it difficult to generate images of different scales

If you want to generate images of multiple scales, you need to change the network structure and retrain, which greatly increases the consumption of computing resources

Images