.png){kind=icon}
.png){kind=icon}
Return
Target Imaging Based on Implicit Interferometric Radiometer
.png){kind=icon}
PRE
AI
DOI：10.1109/TAES.2025.3637787.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Target imaging is a crucial component of target detection and recognition applications in synthetic aperture interferometric radiometer (SAIR). Traditional target imaging frameworks are usually based on hand-crafted priors or deep learning (DL) techniques. The former improves image quality by introducing prior information, but has scene dependence and limited performance. The latter learns the inverse mapping from visibility samples to target images, but relies on large labeled datasets and exhibits significant task specificity. In this article, we propose a target imaging framework, namely implicit interferometric radiometer (IIR). It employs the implicit neural network prior (INNP) to perform high-accuracy SAIR target image reconstruction, which offers low impedance to signals and high impedance to noise and degradations. First, an untrained neural network model is constructed with fixed random noise as input, and the network implicit inductive bias is exploited to generate an estimated result of the target image. The estimated brightness temperature (BT) image is then forward-modeled through the system response matrix to obtain corresponding simulated visibility samples. A joint loss function is formulated, consisting of a fidelity term between measured and simulated visibility samples and customized regularization terms, to guide the network training. The untrained network weights are iteratively optimized by backpropagation to finally recover the desired target image. The proposed target imaging framework operates without reliance on training data and has good generalization capability. The feasibility and effectiveness of IIR are verified using both simulated and real data. The results indicate that IIR achieves better performance compared with conventional methods in terms of image quality, noise suppression, and robustness against missing baselines.
Keywords:
Implicit interferometric radiometer (IIR)
synthetic aperture interferometric radiometer (SAIR)
target imaging
untrained neural network
Journal
IF:
5.7
Papers:
686
Citations:
2.4W