MSTAR SAR-ATR 을 위한 비자기지도 기반 표적 정보 보존 및 SAR speckle 제거 AI 모델 연구

다국어 초록 (Multilingual Abstract)

However, due to the coherent imaging property of SAR, speckle noise inevitably
appears in SAR images. This speckle noise distorts important target-related
structures such as object contours, scattering points, and shadows, which leads to
degraded ATR performance. Recently, deep learning-based despeckling methods
using convolutional neural networks (CNNs) or transformers have shown
performance.
However, most deep learning-based methods utilize pair data of clean and noisy
images, and SAR images are very difficult to construct clean data without noise.
Therefore, when applying existing unsupervised despeckling models to the Moving
and Stationary Target Acquisition and Recognition (MSTAR) SAR-ATR task,
problems such as loss of contour of the target, reduction of scattering point, and
shadow distortion often occur.
To overcome these limitations, we propose an unsupervised target information
retention SAR speckle reduction model for MSTAR SAR-ATR, which considers the
target's information retention together while suppressing the speckle.
Considering that the target is mostly distributed in the center of the image, a
centralized perceptual spatial attention (CASA) module is introduced into the
network. In addition, the mask guide learning strategy is applied to learn by
separating the statistical and structural characteristics of the object, shadow, and
background from the MSTAR image. Several visual comparisons and representative
speckle statistical indicators such as equivalent appearance coefficient (ENL),
speckle index, and edge conservation measurements were used for the proposed
model performance evaluation.
As a result, it was confirmed that the proposed approach reduces speckle and
preserves the target structure better than the existing methods.

목차 (Table of Contents)

• 1.서 론 11
• 1.1 연구 배경 11
• 1.2 MSTAR SAR-ATR 환경의 특수성 13
• 1.3 연구 목적 및 접근 개요 13
• 2. 이론적 배경 15
• 2.1 SAR 영상 형성 원리 15
• 2.2 SAR Speckle Noise 의 발생 메커니즘 17
• 2.3 Multiplicative Noise 모델 17
• 2.4 SAR Speckle 의 통계적 특성 18
• 3. SAR Despeckling 기법 21
• 3.1 전통적 기법 21
• 3.2 변환 영역 기반 기법 22
• 3.3 딥러닝 기반 SAR 기법 23
• 3.4 기존 딥러닝 기반 기법의 한계 25
• 4. 문제 정의 및 연구 접근 전략 26
• 4.1 문제 정의 26
• 4.2 연구 접근 전략 26
• 5. 제안 기법 29
• 5.1 전체 프레임워크 개요 29
• 5.2 Center-Aware Spatial Attention 기반 구조 설계 30
• 5.3 Mask-guided 디스페클링 31
• 6 실험 및 결과 분석 33
• 6.1 실험 데이터셋 및 환경 설정 33
• 6.2 비교 방법 및 실험 구성
• 6.3 정량적 평가 지표 36
• 6.3.1 ENL (Equivalent Number of Looks) 36
• 6.3.2 Speckle Index (SI) 36
• 6.3.3 Edge Preservation 지표 37
• 6.4 정성적 평가 지표 38
• 6.5 결과 및 비교 분석 39
• 6.6 결과 해석 41
• 7. 결론 및 향후 연구 42
• 7.1 연구 결과 요약 42
• 7.2 연구의 의의 및 한계 42
• 참고문헌 45
• Abstract 48