SDF-Net: Structure-Aware Disentangled Feature Learning for Opticall-SAR Ship Re-identification
Authors: Furui Chen, Han Wang, Yuhan Sun, Jianing You, Yixuan Lv, Zhuang Zhou, Hong Tan, Shengyang Li
Organizations: Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences, Beijing 100094, China · Key Laboratory of Space Utilization, Chinese Academy of Sciences, Beijing 100094, China · University of Chinese Academy of Sciences, Beijing 100049, China · School of Software, Beihang University, Beijing 100191, China
Cross-modal ship re-identification (ReID) between optical and synthetic aperture radar (SAR) imagery is fundamentally challenged by the severe radiometric discrepancy between passive optical imaging and coherent active radar sensing. While existing approaches primarily rely on statistical distribution alignment or semantic matching, they often overlook a critical physical prior: ships are rigid objects whose geometric structures remain stable across sensing modalities, whereas texture appearance is highly modality-dependent. In this work, we propose SDF-Net, a Structure-Aware Disentangled Feature Learning Network that systematically incorporates geometric consistency into optical--SAR ship ReID. Built upon a ViT backbone, SDF-Net introduces a structure consistency constraint that extracts scale-invariant gradient energy statistics from intermediate layers to robustly anchor representations against radiometric variations. At the terminal stage, SDF-Net disentangles the learned representations into modality-invariant identity features and modality-specific characteristics. These decoupled cues are then integrated through a parameter-free additive residual fusion, effectively enhancing discriminative power. Extensive experiments on the HOSS-ReID dataset demonstrate that SDF-Net consistently outperforms existing state-of-the-art methods. The code and trained models are publicly available at https://github.com/cfrfree/SDF-Net.
Figures & tables
Fig. 1: Illustration of the optical–SAR ship re-identification task. The framework aims to match the same ship identity across vastly different optical and SAR sensor modalities.
Fig. 2: Architectural pipeline of the proposed SDF-Net. The framework builds upon the TransOSS base architecture (cross-modal dual-head tokenizer, MIE, SSE, and ViT-B/16 backbone), initialized with weights pre-trained on optical–SAR image pairs from the TransOSS framework. SDF-Net introduces two novel physics-guided modules: (b) Structure-Aware Consistency Learning (SCL), which extracts and aligns intermediate gradient energy to anchor representations on modality-invariant structural primitives; and (c) Disentangled Feature Learning (DFL), which decouples shared identity features from modality-specific features and fuses them via additive residual refinement. In the ablation study, the baseline disables both SCL and DFL, providing a controlled comparison under the same hyperparameter configuration.
Protocol
Modality
Query images
Gallery images
Total
All-to-All
Optical
88
403
491
SAR
88
190
278
Subtotal
176
593
769
Optical-to-SAR
Optical
65
0
65
SAR
0
190
190
Subtotal
65
190
255
TABLE I: Detailed statistical distribution of the HOSS-ReID testing set across different evaluation protocols.
Task Type
Method
Venue
All-to-All
Optical-to-SAR
SAR-to-Optical
mAP
R1
R5
R10
mAP
R1
R5
R10
mAP
R1
R5
R10
General model
ViT-base [ 29 ]
Arxiv2020
43.0
56.2
64.8
69.9
21.5
12.3
33.8
55.4
17.9
10.4
25.4
32.8
DeiT-base [ 30 ]
ICML2021
47.2
58.1
69.6
74.1
25.9
16.1
36.7
59.1
26.1
10.3
35.7
52.0
Single modality ReID
AGW [ 8 ]
TPAMI2021
43.6
57.4
64.2
68.8
17.2
7.7
29.2
38.5
21.1
14.9
34.3
46.3
TransReID [ 10 ]
ICCV2021
48.1
60.8
69.3
73.9
27.3
18.5
40.0
58.5
20.9
11.9
34.3
43.3
SOLIDER [ 31 ]
CVPR2023
38.2
50.6
63.1
69.9
23.1
12.3
38.5
52.3
14.6
10.4
16.4
31.3
TABLE II: Quantitative evaluation of the proposed SDF-Net compared with state-of-the-art methods on the HOSS-ReID benchmark. All values are reported in percentage (%).
SCL
DFL
All-to-All
Optical-to-SAR
SAR-to-Optical
mAP
R1
mAP
R1
mAP
R1
✗
✗
58.6
67.6
46.5
32.3
44.5
38.8
✓
✗
59.2
66.5
47.6
32.3
46.6
37.3
✗
✓
59.8
69.9
49.3
35.4
41.4
31.3
✓
✓
60.9
69.9
50.0
35.4
46.6
38.8
TABLE III: Ablation study highlighting the effectiveness of the SCL and DFL modules on the HOSS-ReID benchmark. All values are reported in percentage (%).
Bs
All-to-All
Optical-to-SAR
SAR-to-Optical
mAP
R1
mAP
R1
mAP
R1
2
59.7
68.2
49.0
35.4
46.0
40.3
4
60.4
68.8
50.5
38.5
45.3
38.8
6
60.9
69.9
50.0
35.4
46.6
38.8
8
58.4
65.3
48.7
35.4
45.5
40.3
10
58.7
66.5
48.9
33.8
44.7
37.3
TABLE IV: Ablation study on the structural extraction layer index Bs . The results indicate that extracting structural priors at intermediate layers such as Bs=6 , achieves the optimal trade-off between spatial fidelity and semantic abstraction. All values are reported in percentage (%).
Fusion Strategy
All-to-All
Optical-to-SAR
SAR-to-Optical
mAP
R1
mAP
R1
mAP
R1
Specific-only ( fsp )
58.7
67.6
44.3
29.2
43.9
35.8
Shared-only ( fsh )
59.2
68.2
49.8
33.8
43.1
37.3
Concatenation
59.5
68.8
50.7
35.4
45.1
35.8
Additive (Ours)
60.9
69.9
50.0
35.4
46.6
38.8
TABLE V: Ablation study on feature fusion strategies. The additive interaction acts as a parameter-efficient residual refinement, outperforming concatenation and single-branch baselines. All values are reported in percentage (%).
Method
Params (M)
FLOPs (G)
mAP (%)
Rank-1 (%)
TransOSS
86.24
22.25
57.4
65.9
SDF-Net
86.24
22.42
60.9
69.9
TABLE VI: Computational complexity and performance comparison between the baseline and the proposed SDF-Net.
Fig. 3: Hyper-parameter sensitivity analysis of SDF-Net. The heatmaps illustrate the variation of mAP, on the left, and Rank-1, on the right in percentage (%), under different combinations of the orthogonality constraint weight λorth and the structure consistency weight λstruct .
Scale
Baseline
SCL-only
DFL-only
SDF-Net
mAP
R1
mAP
R1
mAP
R1
mAP
R1
0.25 ×
53.2
63.1
53.8
61.9
52.4
62.5
53.1
62.5
0.50 ×
56.0
64.2
57.3
64.2
57.1
67.0
57.4
65.9
0.75 ×
57.5
64.8
57.3
64.2
58.8
68.8
58.8
67.0
1.00 × (native)
58.6
67.6
59.2
66.5
59.8
69.9
60.9
69.9
1.50 ×
57.8
66.5
58.0
64.8
58.9
68.2
59.6
67.6
TABLE VII: Resolution stability with ablation comparison. SCL-only = baseline + SCL; DFL-only = baseline + DFL; SDF-Net = SCL + DFL. Bold entries mark the best value in each row. All metrics in percentage (%).
θ
Baseline
SCL-only
DFL-only
SDF-Net
mAP
R1
mAP
R1
mAP
R1
mAP
R1
5∘
11.4
15.3
13.3
19.9
13.1
18.8
12.2
18.2
15∘
31.5
43.8
31.2
42.0
31.4
44.9
32.6
45.5
25∘
41.7
52.8
41.9
55.1
42.3
54.5
42.6
55.1
35∘
46.3
56.8
45.2
55.7
46.2
57.4
47.1
58.5
TABLE VIII: Robustness to SAR projective distortion—foreshortening and layover—with ablation comparison. θ denotes the incidence angle; shallower angles produce more severe distortion. All metrics in percentage (%). Models are evaluated without geometric augmentation training.
Blur (px)
Baseline
SCL-only
DFL-only
SDF-Net
mAP
R1
mAP
R1
mAP
R1
mAP
R1
0 (native)
58.6
67.6
59.2
66.5
59.8
69.9
60.9
69.9
3
58.1
66.5
59.0
65.9
58.8
67.6
59.7
67.6
5
57.6
65.3
57.8
64.8
57.5
66.5
58.4
67.0
7
57.4
66.5
57.5
65.3
56.4
66.5
57.0
65.3
9
55.8
64.8
56.8
65.9
55.7
65.9
55.9
65.3
TABLE IX: Robustness to simulated SAR azimuthal defocusing, or motion blur, with ablation comparison. All metrics in percentage (%). SCL-only remains the top performer under strong blur.
Fig. 4: Grad-CAM visualization of the spatial attention maps generated by SDF-Net. From left to right within each group: the input optical image, the corresponding optical attention map, the SAR attention map, and the input SAR image. The network consistently focuses on the modality-invariant ship hull structure, effectively suppressing optical sea clutter and penetrating SAR speckle noise.
Fig. 5: Visual evolution of feature heatmaps across different Transformer layers. For each modality group, the columns from left to right represent the original input image followed by the feature activation maps extracted from layers 2, 4, 6, 8, 10, and 12. Intermediate representations at layer 6 successfully isolate the modality-invariant geometric structure, whereas shallow layers are corrupted by sensor noise and deep layers suffer from spatial semantic collapse.
Fig. 6: Qualitative comparison of retrieval results between the baseline and the proposed SDF-Net across three evaluation protocols. The left, middle, and right columns correspond to the All-to-All, Optical-to-SAR, and SAR-to-Optical settings, respectively. Green bounding boxes indicate correct matches, while red bounding boxes represent incorrect matches. SDF-Net exhibits superior robustness against modality-specific noise, consistently retrieving correct physical identities at higher ranks compared to the appearance-biased baseline.
Fig. 7: Failure case analysis of SDF-Net. Left panel: SAR-to-Optical retrieval. Right panel: Optical-to-SAR retrieval. Each row shows the query, the ground-truth match, and the incorrect top-1 prediction. Failures are predominantly caused by geometrically similar ship structures and low-resolution targets with weak structural cues.
School of Artificial Intelligence, Beijing Normal University · Beijing Key Laboratory of Artificial Intelligence for Education · Engineering Research Center of Intelligent Technology and Educational Application, Ministry of Education