High-resolution Imaging Method for Through-the-wall Radar Based on Transfer Learning with Simulation Samples

Abstract

This paper addresses the problem of high-resolution imaging of shadowed multiple-targets with limited labeled data, by proposing a transfer-learning-based method for through-the-wall radar imaging. First, a generative adversarial sub-network is developed to facilitate the migration of labeled simulation data to measured data, overcoming the difficulty of generating labeled data. This method incorporates an attention mechanism, adaptive residual blocks, and a multi-scale discriminator to improve the quality of image migration. It also incorporates a structural consistency loss function to minimize perceptual differences between images. Finally, the labeled data are used to train the through-the-wall radar target-imaging sub-network, achieving high-resolution imaging of multiple targets through walls. Experimental results show that the proposed method effectively reduces discrepancies between simulated and obtained images, and generates pseudo-measured images with labels. It systematically addresses issues such as side/grating ghost interference, target image defocusing, and multi-target mutual interference, significantly improving the multi-target imaging quality of the through-the-wall radar. The imaging accuracy achieved is 98.24%, 90.97% and 55.17% for single, double, and triple-target scenarios, respectively. Compared with CycleGAN, the imaging accuracy for the corresponding scenarios is improved by 2.29%, 40.28% and 15.51%, respectively.

Keywords

• transfer learning
• generative adversarial nets (gan)
• domain adaptation
• through-the-wall radar (twr)
• high-resolution imaging