논문 상세보기
Road Segmentation-Aware SAR Image Enhancement Using Task-Driven Generative Adversarial Networks (TD-GANs) KCI 등재
과업 기반 생성적 적대 신경망(TD-GAN)을 이용한 도로 분할 인지형 SAR 영상 향상 기초연구
- 언어ENG
- URLhttps://db.koreascholar.com/Article/Detail/453290
pp.25-37
Satellite synthetic aperture radar (SAR) enables all-weather road monitoring; however, multiplicative speckle noise degrades the thin linear returns on which road segmentation networks rely. At 10 m resolution in the Gaofen-3 (GF-3) fine stripmap II mode, road widths span one to two pixels while conventional enhancement optimized for pixel fidelity either fails to improve or degrades segmentation. Thus, this study proposes a task-driven generative adversarial network (TD-GAN), which is a 0.34 M-parameter Swin Transformer generator whose speckle-aware stripe attention block (SSAB) downweights speckle regions while preserving road structures. The generator is trained on unpaired 10 m/3 m data via a CycleGAN and then fine-tuned by backpropagating the loss of a frozen 3 m-trained segmenter. On the GF-3 SAR road benchmark, TD-GAN increased the mean IoU of a U-Net segmenter from 0.700 to 0.799 (+0.099) and outperforms the 11.37 M ResNet CycleGAN baseline by +0.023 mean IoU using 33× fewer generator parameters. Of the +0.099 improvement, +0.090 is attributable to the task-driven fine-tuning paradigm established in prior work, while the remaining +0.009 is the net contribution of the proposed SSAB. The primary value of the SSAB lies in its consistency instead of its magnitude: it consistently yields positive gains across all six evaluated segmentation backbones (paired t(5)=10.52, p<0.001, 95% CI [+0.0066, +0.0109]) and in 25 of 30 cross-backbone transfer configurations. Across six backbones under a matched 100-epoch protocol, TD-GAN recovers 39%–67% of the segmentation accuracy gap between the 10m baseline and the 3 m oracle without requiring paired supervision. Furthermore, when a single generator is applied directly to a segmenter without retraining or fine-tuning, it improves all 30 cross-backbone configurations over the native 10 m input, retaining a median of 58% of the backbone-matched gain. Thus, TD-GAN serves as a lightweight preprocessing module highly suitable for SAR-based road monitoring in disaster response and continental-scale mapping.
1. Introductuon
2. Related Work
3. Proposed Method: TD-GAN
4. Experiments
4.1 Experimental setup
4.2 Results
5. Discussion
6. Conclusions and Future Studies
REFERENCES
