논문 상세보기
Underwater Sonar Image Classification based on Vision Transformer with Metric Learning KCI 등재
비전 트랜스포머 및 메트릭 러닝 기반 수중 소나 이미지 분류
- 언어KOR
- URLhttps://db.koreascholar.com/Article/Detail/449284
pp.155-164
Underwater sonar image classification is essential for maritime surveillance, autonomous navigation, and underwater target identification, where optical sensing is often restricted by turbidity and light attenuation. To enhance the robustness of sonar-based perception under such challenging conditions, this study proposes a metric-enhanced Vision Transformer (ViT) framework that integrates Siamese-based representation alignment with distance-regularized classification. In the first stage, a Siamese pre-training strategy is employed to align embeddings of positive pairs, encouraging directionally consistent representations that improve class separability even under severe noise and viewpoint variations. In the second stage, the pretrained ViT encoder is frozen, and five classifiers—Linear, Cosine, Proxy, and their Mahalanobis-regularized variants—are systematically evaluated to investigate the effect of embedding normalization and distributional alignment. Experimental results on the UATD dataset demonstrate that the Siamese-trained ViT produces more stable and discriminative features than both ResNet-50 and standard ViT-S. Among the classifiers, the Mahalanobis-regularized cosine classifier achieves the highest, showing significant reductions in misclassification between visually similar classes such as cube and square cage. Overall, the proposed approach highlights the effectiveness of combining ViT with metric learning and covariance-aware distance normalization for underwater sonar image recognition. The results suggest that metric-enhanced transformers offer a robust and generalizable foundation for sonar-based perception in real maritime environments.
2. 연구배경 및 관련 연구
2.1 수중 소나 이미지 분석
2.2 기존 심층 신경망 기반 접근법
2.3 전이 학습
2.4 메트릭 러닝
2.5 거리 기반 분류기
2.6 관련연구
3. 제안 방법
3.1 개요
3.2 1단계: 시암 ViT 표현 학습
3.3 2단계: 분류기 학습
4. 실험 및 결과분석
4.1 데이터 셋
4.2 실험 구성
4.3 실험 결과 및 분석
5. 결 론
Acknowledgement
References
