Kyung-Tae Kim

Synthetic aperture radar (SAR) super-resolution aims to overcome the intrinsic resolution limits imposed by system bandwidth and synthetic-aperture length. Existing deep learning-based methods, however, typically rely on naive down-sampling or single-resolution training. Such settings do not adequately reflect SAR-specific resolution properties and often generalize poorly across input resolutions. This paper presents a SAR-specific multi-resolution data augmentation strategy based on frequency-domain band cropping. By controlling bandwidth reduction in range and azimuth while preserving the original image grid, the proposed method generates physics-consistent low-resolution/high-resolution pairs from each original SAR chip. In this sense, data augmentation expands the training distribution along the resolution axis and allows a single super-resolution model to learn across heterogeneous resolution gaps. Experiments on the moving and stationary target acquisition and recognition (MSTAR) dataset show improved PSF width, PSLR/ISLR, PSNR, and SSIM over naive down-sampling under a fair comparison based on matched effective-resolution loss. The multi-resolution strategy is also more robust than mismatched single-resolution training while remaining close to matched single-resolution training at the tested degradation level. Comparisons with RELAX- and CS-based baselines, additional object-level validation on ATRNet, and held-out-scene validation on Capella further support the practical relevance of the framework. Phase analysis indicates that the present gains are strongest in object-level amplitude-domain restoration.

한국전자파학회논문지 (The Journal of Korean Institute of Electromagnetic Engineering and Science)

한국전자파학회논문지 (The Journal of Korean Institute of Electromagnetic Engineering and Science)

한국전자파학회논문지 (The Journal of Korean Institute of Electromagnetic Engineering and Science)

High-resolution inverse synthetic aperture radar (ISAR) imaging is indispensable for target classification but is often compromised by complex 3D rotational motion (RM). Conventional frame selection schemes identify 2D-RM dominant intervals and discard the remaining data, which limits the coherent integration time and significantly degrades Doppler resolution. To overcome this limitation, this letter proposes a soft-integration framework that adaptively incorporates all available data. The proposed method assigns frame-level reliability weights via joint time-frequency analysis to quantify the stability of each interval. By utilizing these weights in a robust parameter estimation process, the framework preserves latent 2D-RM information from 3D-RM dominant intervals while suppressing disruptive out-of-plane effects. Experimental results with simulated and measured datasets demonstrate significant improvements in Doppler resolution and image quality compared to traditional hard-selection methods.

Small drone-borne W-band synthetic aperture radar (SAR) systems are highly susceptible to motion errors that conventional navigation sensors and phase-based autofocus algorithms cannot effectively resolve due to phase wrapping. This paper presents a sensor-independent imaging framework to robustly suppress these errors. First, joint time-frequency analysis is employed to identify and discard motion-corrupted pulses. Subsequently, a compressed sensing-based backprojection algorithm reconstructs high-resolution images from the remaining sparse dataset. To alleviate the substantial memory burden of matrix-based compressed sensing, the reconstruction is reformulated iteratively. Experimental results confirm that the proposed method maintains structural integrity even when up to 60% of the received pulses are corrupted and demonstrates robust focusing down to an SNR of −25 dB. This approach provides a practical, memory-efficient, and cost-effective solution for SAR platforms.

Unsupervised change detection using synthetic aperture radar (SAR) imagery aims to identify changes in terrain or objects by analyzing difference images generated from multi-temporal SAR pairs. Although traditional threshold-based approaches are effective for large-scale changes, they often fail to detect small-scale changes, such as changes in vehicle positions, due to the dominance of unchanged regions in the difference image distribution. To address this issue, we propose a novel fusion-based approach designed to enhance the detectability of small objects by fusing a log-ratio image with a non-coherent change detection image using Berger's method (NCB). A fuzzification process aligns the dynamic ranges of the two images, and an adaptive fuzzy membership function derived from the statistical distribution of the log-ratio image further improves robustness. The fused difference image is generated via the Hadamard product, which reduces clutter-induced false alarms. Experimental results on both simulated and real SAR data sets demonstrate that the proposed method provides a balanced operating threshold on the fused image and achieves the best F1-score on most evaluated scenes.

Inverse synthetic aperture radar (ISAR) imaging is useful for classifying targets of interest. Therefore, it is critical to enhance Doppler resolution, which depends on the uncontrollable target’s motion. Traditional range-Doppler processing can generally compensate for translational motion (TM) and two-dimensional rotational motion (2D-RM) to focus ISAR images. However, it is nearly impossible to compensate for the three-dimensional rotational motion (3D-RM) of highly maneuvering targets. To solve this problem, frame selection approaches have been devised to retain frames with 2D-RM only, while discarding frames with 3D-RM. However, conventional frame selection methods usually have limited Doppler resolution against highly maneuvering targets, because each selected time frame with 2D-RM is often very short. This paper proposes an imaging framework to overcome the aforementioned 3D-RM problems, by exploiting frame selection and sparse-aperture (SA)-ISAR imaging simultaneously. The proposed approach first identifies time frames with dominant 2D-RM. Notably, time frames with 2D-RM consist of non-uniform samples along the slow time. This problem poses challenges for adopting conventional RM compensation methods. The proposed framework compensates for the nonlinear 2D-RM on non-uniform grid by employing dictionary estimation using the parametric signal reconstruction (DEPSR). Experimental results for both the simulated and real measurement datasets demonstrate the effectiveness of the proposed framework.