SAR image sample generation method based on sketch and structure generation countermeasure network

Abstract

The invention provides a SAR image sample generation method based on a sketch and structure generation antagonistic network, which mainly solves the problem of sample imbalance in the semantic segmentation of the SAR image. The method comprises the following steps of: (1) sketching a SAR image to obtain a sketch map; (2) according to the region map of SAR image, extracting the small extremely heterogeneous region; (3) constructing a paired data set of sketch blocks-SAR image blocks; (4) selecting the samples from the data set to compose the training set and the test set; (5) constructing a generating antagonistic network based on sketch information and structural constraints; (6) training the sketch fitting network, the discrimination network and the generation network alternately by sketch line loss, antagonistic loss and generator loss,; (7), inputting a test sketch block to the trained generation network to obtain a generated SAR image block. The invention can generate SAR image samples coinciding with the original SAR image ground object structure according to the sketch map, and can solve the sample unbalance problem of extremely uneven region classification of the SAR image.

Description

technical field [0001] The invention belongs to the technical field of image generation, and further relates to deep learning, an image generation method for generating confrontation network GANs (Generative Adversarial Networks) and synthetic aperture radar SAR (Synthetic ApertureRadar) images. Background technique [0002] Sample imbalance is a common and unavoidable problem in SAR image classification and segmentation. Commonly used machine learning models usually assume that the samples are unbiased samples of the real distribution. However, in image classification, due to the disparity between rare and rich samples, most models have extremely poor recall rates for rare categories on the test set. Similarly, the existence of small areas puts forward higher requirements on the segmentation technology of SAR images. If image segmentation is regarded as a classification problem at the image pixel level, the number of samples in small areas is far less than that in larger ar...

AI Technical Summary

Problems solved by technology

[0007] The purpose of the present invention is to provide a SAR image sample generation method based on sketch and structure generation confrontation network, which solves the existing sample imbalance, which leads to the defect that the SAR image cannot be classified and segmented

Images