Paper deep dive
Deformation-Free Cross-Domain Image Registration via Position-Encoded Temporal Attention
Yiwen Wang, Jiahao Qin
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 7/21/2026, 2:35:37 AM
Summary
The paper introduces GPEReg-Net, a deformation-free cross-domain image registration framework that decomposes images into domain-invariant scene representations and global appearance statistics. By recombining these via Adaptive Instance Normalization (AdaIN) and utilizing a position-encoded temporal attention mechanism for sequential frames, the method eliminates explicit deformation field estimation. It achieves state-of-the-art performance on FIRE-Reg-256 and HPatches-Reg-256 benchmarks, outperforming traditional and deep learning baselines in both accuracy and inference speed.
Entities (10)
Relation Signals (7)
GPEReg-Net → achievesmetric → PSNR 33.47 dB
confidence 95% · FIRE: SSIM = 0.928, PSNR = 33.47 dB
GPEReg-Net → achievesmetric → SSIM 0.928
confidence 95% · FIRE: SSIM = 0.928
GPEReg-Net → evaluatedon → HPatches-Reg-256
confidence 95% · and HPatches-Reg-256
GPEReg-Net → evaluatedon → FIRE-Reg-256
confidence 95% · evaluate on two benchmarks: FIRE-Reg-256
GPEReg-Net → uses → AdaIN
confidence 95% · registration reduces to recombining the scene structure of the moving image with the appearance of the fixed image via Adaptive Instance Normalization (AdaIN)
GPEReg-Net → usesmodule → Position-Encoded Temporal Attention
confidence 92% · introduce a position-encoded cross-frame attention mechanism
GPEReg-Net → outperforms → SAS-Net
confidence 90% · running 1.87x faster than SAS-Net
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We address the problem of cross-domain image registration, where paired images exhibit coupled geometric misalignment and domain-specific appearance shift. We formalize this as a factorization problem: decomposing each image into a domain-invariant scene representation and a global appearance statistic, such that registration reduces to recombining the scene structure of the moving image with the appearance of the fixed image via Adaptive Instance Normalization (AdaIN). This factorization eliminates the need for explicit deformation field estimation. To exploit temporal coherence in sequential acquisitions, we introduce a position-encoded cross-frame attention mechanism that fuses learnable and sinusoidal position embeddings with multi-head attention over a sliding window of neighboring frames, enriching the scene representation with inter-frame context. We instantiate this framework as GPEReg-Net and evaluate on two benchmarks: FIRE-Reg-256 (retinal fundus, semi-rigid) and HPatches-Reg-256 (synthetic textured patches, affine). GPEReg-Net achieves state-of-the-art performance on both benchmarks (FIRE: SSIM = 0.928, PSNR = 33.47 dB; HPatches: SSIM = 0.450, PSNR = 21.01 dB), surpassing all baselines, including deformation-based methods, while running 1.87x faster than SAS-Net. Code: this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2602.15959v2
- Canonical: https://arxiv.org/abs/2602.15959v2
Trouble viewing inline? Open PDF directly →
Full Text
24,348 characters extracted from source content.
Expand or collapse full text
Deformation-Free Cross-Domain Image Registration via Position-Encoded Temporal Attention Yiwen Wang Jiahao Qin Corresponding author: jiahao.qin19@gmail.com Abstract We address the problem of cross-domain image registration, where paired images ImI_m and IfI_f exhibit coupled geometric misalignment and domain-specific appearance shift. We formalize this as a factorization problem: decomposing each image into a domain-invariant scene representation ∈ℝC×H×W s ^C× H× W and a global appearance statistic ∈ℝd a ^d, such that registration reduces to recombining the scene structure of ImI_m with the appearance of IfI_f via Adaptive Instance Normalization (AdaIN). This factorization eliminates the need for explicit deformation field estimation. To exploit temporal coherence in sequential acquisitions, we introduce a position-encoded cross-frame attention mechanism that fuses learnable and sinusoidal position embeddings with multi-head attention over a sliding window of k neighboring frames, enriching the scene representation with inter-frame context. We instantiate this framework as GPEReg-Net and evaluate on two benchmarks: FIRE-Reg-256 (retinal fundus, semi-rigid) and HPatches-Reg-256 (synthetic textured patches, affine). GPEReg-Net achieves state-of-the-art performance on both benchmarks (FIRE: SSIM = 0.928, PSNR = 33.47 dB; HPatches: SSIM = 0.450, PSNR = 21.01 dB), surpassing all baselines including deformation-based methods, while running 1.87×1.87× faster than SAS-Net. Code: https://github.com/JiahaoQin/GPEReg-Net. 1 Introduction Let Im,If∈ℝH×WI_m,I_f ^H× W denote a moving image and a fixed image, respectively. Image registration seeks a mapping :Im↦I^rT:I_m I_r such that I^r≈If I_r≈ I_f under a suitable similarity measure [25, 5]. In cross-domain settings—where ImI_m and IfI_f are drawn from different intensity distributions due to varying acquisition conditions (e.g., subject motion in retinal imaging [8], viewpoint changes in natural images [3])—the brightness constancy assumption Im()≈If(+)I_m( x)≈ I_f( x+ u) underlying conventional methods [5] is systematically violated. Limitations of existing approaches. Classical methods—SIFT [12], Demons [27], optical flow [9], SyN [1]—estimate spatial correspondences directly but degrade when intensity distributions differ. Learned deformation methods (VoxelMorph [2], TransMorph [4], spatial transformer networks [11], Laplacian pyramid approaches [13]) parameterize T as a displacement field () u( x) but inherit the same intensity assumption. Scene-appearance separation frameworks [23, 21] address domain shift through disentangled latent spaces but employ complex generative architectures without temporal awareness. Progressive strategies [22] improve alignment iteratively but do not explicitly factorize appearance from scene content. Key insight. We observe that the cross-domain registration problem admits a natural factorization: each image can be decomposed into a domain-invariant scene representation s (encoding spatial structure) and a domain-specific appearance statistic a (encoding intensity profile). Registration then reduces to recombining m s_m with f a_f via Adaptive Instance Normalization (AdaIN) [10], yielding I^r=(AdaIN(m,f)) I_r=D(AdaIN( s_m, a_f)) without estimating any deformation field [14, 16, 18]. Furthermore, in sequential acquisitions, temporal coherence between consecutive frames can be exploited via position-encoded cross-frame attention [26, 17, 15] to improve inter-frame consistency. We instantiate this insight as GPEReg-Net and make the following contributions: 1. Scene-appearance factorization. We formalize cross-domain registration as a latent factorization problem and propose an encoder-decoder architecture that decomposes images into domain-invariant scene features ∈ℝ64×H×W s ^64× H× W and global appearance codes ∈ℝ32 a ^32, with AdaIN-based recombination that eliminates deformation field estimation entirely. 2. Position-encoded temporal attention. We introduce a Global Position Encoding (GPE) module that fuses learnable position embeddings, sinusoidal encoding, and multi-head cross-frame attention over a sliding window of k neighbors, enabling the model to exploit temporal structure in sequential acquisitions. 3. Comprehensive cross-domain evaluation. We validate on two diverse benchmarks—FIRE-Reg-256 (retinal fundus, semi-rigid) and HPatches-Reg-256 (synthetic patches, affine)—achieving state-of-the-art results on both while maintaining real-time throughput (6969 FPS). 2 Proposed Method: GPEReg-Net 2.1 Problem Formulation and Architecture Overview Given a moving image Im∈ℝH×WI_m ^H× W and a fixed image If∈ℝH×WI_f ^H× W drawn from potentially different intensity distributions, our goal is to learn a mapping θ:(Im,If)↦I^rT_θ:(I_m,I_f) I_r such that I^r I_r is geometrically aligned with IfI_f while adopting its intensity profile, without estimating an explicit deformation field :ℝ2→ℝ2 u:R^2 ^2. We decompose θT_θ into four learned modules (fig. 1): 1. SceneEncoder S: Extracts domain-invariant scene features =(Im) s=S(I_m) via instance normalization, discarding appearance information. 2. AppearanceEncoder A: Extracts a global appearance code =(If) a=A(I_f) that captures the intensity profile of the target domain. 3. Global Position Encoding G: Enhances scene features with temporal context: ~=(,t) s=G( s,t), where t is the frame index. 4. ImageDecoder D: Reconstructs the registered output using AdaIN modulation: I^r=(~,) I_r=D( s, a). Figure 1: (a) Registration example on HPatches-Reg-256 [3]: the fixed image IfI_f, moving image ImI_m with geometric misalignment, the registered output I^r I_r produced by GPEReg-Net, and the absolute difference |I^r−If|| I_r-I_f|. A dashed arrow indicates the appearance pathway from IfI_f. (b) Overview of GPEReg-Net. The SceneEncoder S extracts domain-invariant features s from the moving image ImI_m using instance normalization. The AppearanceEncoder A extracts a 32-dim global appearance code a from the fixed image IfI_f. The GPE module G enriches scene features with temporal context via learnable position embeddings, sinusoidal encoding, and cross-frame attention, producing ~ s. The ImageDecoder D uses AdaIN to modulate enhanced scene features with the target appearance, producing the registered output I^r I_r without any explicit deformation field. 2.2 Scene-Appearance Factorization via Dual Encoders The SceneEncoder extracts domain-invariant features from the moving image using a U-Net [24] backbone with residual connections [7] and Instance Normalization (IN), which strips per-instance intensity statistics while preserving spatial structure. It produces a 64-dimensional scene feature map ∈ℝ64×H×W s ^64× H× W via four encoding levels (Cl∈32,64,128,256C_l∈\32,64,128,256\) with bilinear upsampling decoder and skip connections. The AppearanceEncoder extracts a 32-dimensional global appearance code ∈ℝ32 a ^32 from the fixed image via four stride-2 convolutions, Global Average Pooling, and two FC layers, capturing target domain intensity statistics without spatial information. The ImageDecoder reconstructs the registered output by injecting the target appearance into scene features using Adaptive Instance Normalization (AdaIN) [10]: AdaIN(,)=γ()⋅−μ()σ()+β(),AdaIN( s, a)=γ( a)· s-μ( s)σ( s)+β( a), (1) where γ()γ( a) and β()β( a) are learned affine parameters. Three AdaIN-Conv blocks (64, 32, 16 channels) progressively reconstruct the output I^r I_r without any spatial deformation field. 2.3 Position-Encoded Cross-Frame Temporal Attention The GPE module exploits temporal structure in sequential acquisitions by enriching scene features with inter-frame positional context before decoding. For frame index t∈0,…,N−1t∈\0,…,N-1\, the module computes a position-aware representation by combining: (1) a learnable position embedding t∈ℝ64 e_t ^64 from an embedding table ∈ℝN×64 E ^N× 64; (2) a fixed sinusoidal encoding t∈ℝ64 p_t ^64 [26] for smooth frequency-based interpolation; and (3) cross-frame multi-head attention (H=4H\!=\!4) that queries the current frame’s spatially-averaged scene features against a running cache of k=2k\!=\!2 neighboring frames. The concatenated embeddings are fused via a two-layer MLP to produce t g_t, which is added to the scene features with scaling α=0.1α\!=\!0.1: ~=+α⋅Wproj(t+t), s= s+α· W_proj( g_t+ c_t), (2) where t c_t is the cross-frame attention output and WprojW_proj broadcasts to the spatial feature dimensions. 2.4 Training Objective We optimize a bi-objective loss combining pixel-wise reconstruction fidelity with a scene-level factorization regularizer: ℒ=ℒrecon+λ⋅ℒscene,L=L_recon+λ·L_scene, (3) where ℒrecon=‖I^r−If‖1L_recon=\| I_r-I_f\|_1 enforces pixel-level reconstruction fidelity, and ℒscene=‖(Im)−(If)‖22L_scene=\|S(I_m)-S(I_f)\|_2^2 is a factorization consistency term that encourages domain-invariant scene representations: if the factorization is correct, both images should map to the same scene code regardless of their appearance domain. The weighting λ=10.0λ\!=\!10.0 is determined via multi-task balancing [20, 19]. 3 Experiments 3.1 Datasets and Implementation Details Datasets. We evaluate on two benchmarks spanning diverse imaging modalities and deformation types. (1) FIRE-Reg-256 [8]: 134 medical image pairs preprocessed to 256×256256×256 patches (8,018/978/973 train/val/test), featuring semi-rigid deformations from clinical imaging. (2) HPatches-Reg-256 [3]: derived from the HPatches local descriptor benchmark, comprising synthetic textured image patches with Gaussian blob structures and structured noise patterns, with registration pairs generated by random affine transformations (rotation ±15∘± 15 , translation ±20± 20 px, scaling 0.850.85–1.15×1.15×, shear ±10∘± 10 ). The dataset contains 8,000/500/500 train/val/test pairs at 256×256256×256 resolution. Implementation. GPEReg-Net contains 3.40M parameters. Training uses the Adam optimizer (lr=10−4lr=10^-4, weight decay 10−510^-5) with cosine annealing over 30 epochs, batch size 8, and gradient clipping (max norm 1.0). Mixed precision (AMP) is employed for memory efficiency. All experiments use a single NVIDIA RTX 5090 GPU. Evaluation metrics. We report NCC, SSIM [28], and PSNR to quantify alignment quality between the registered output and the fixed image. 3.2 Quantitative Evaluation on FIRE-Reg-256 table 1 evaluates semi-rigid generalization on FIRE-Reg-256 [8] (973 test patches). The unregistered NCC = 0.762 reflects highly overlapping, well-aligned pairs; traditional warping methods degrade performance below this baseline. All methods are trained from scratch on FIRE-Reg-256 for 20 epochs. Table 1: Registration on FIRE-Reg-256 [8] (973 test patches). Our method in bold. Method NCC↑ SSIM↑ PSNR↑ Traditional Methods Unregistered 0.762 0.494 22.36 SIFT [12] 0.449 0.463 16.39 Demons [27] 0.672 0.528 17.45 Optical Flow [9] 0.552 0.506 16.77 SyN [1] 0.549 0.521 15.76 Deep Learning Methods VoxelMorph [2] 0.820 0.916 25.42 TransMorph [4] 0.832 0.876 25.51 SAS-Net [23] 0.748 0.855 32.21 GPEReg-Net (Ours) 0.851 0.928 33.47 GPEReg-Net achieves the highest scores across all three metrics: NCC of 0.851, SSIM of 0.928, and PSNR of 33.47 dB on FIRE-Reg-256, surpassing all traditional and deep learning baselines. The NCC improvement over TransMorph (0.832) demonstrates that the scene-appearance disentanglement captures structural alignment at least as effectively as deformation-based approaches, while the superior SSIM and PSNR (versus VoxelMorph’s 0.916 SSIM and SAS-Net’s 32.21 dB PSNR) confirm the benefit of AdaIN-based appearance transfer [21]. Compared to SAS-Net [23], the GPE module provides additional temporal context that improves consistency across sequential frames, contributing to the 1.26 dB PSNR gain. These results confirm that position-aware scene-appearance disentanglement [19] generalizes effectively across imaging domains. 3.3 Computational Efficiency table 2 compares the computational cost of GPEReg-Net against baseline methods, benchmarked on an NVIDIA RTX 5090 GPU. Table 2: Computational efficiency comparison. Parameters (M), inference latency (ms), and throughput (FPS) benchmarked on an RTX 5090 GPU. Method Params (M) Latency (ms) FPS VoxelMorph [2] 0.10 3.06 327 TransMorph [4] 0.17 2.85 351 SAS-Net [23] 3.35 27.21 37 GPEReg-Net (Ours) 3.40 14.52 69 GPEReg-Net (3.40M params) achieves 69 FPS (14.52 ms latency on RTX 5090), representing a 1.87× speedup over SAS-Net (3.35M params, 37 FPS) due to its simpler AdaIN-based decoding architecture. Deformation-based methods (VoxelMorph 327 FPS, TransMorph 351 FPS) are faster but achieve far lower registration quality. GPEReg-Net’s throughput exceeds typical sequential imaging acquisition rates, enabling real-time processing in research and clinical settings. 3.4 Analysis: Why Does Factorization Work? The effectiveness of the scene-appearance factorization rests on an information-theoretic argument about the complementary roles of the two encoders. Let =(I) s=S(I) and =(I) a=A(I). Instance normalization in S removes per-channel first- and second-order statistics (μc,σc)( _c, _c), retaining only the spatial structure of feature activations. Global average pooling in A discards all spatial information, retaining only channel-wise statistics. These two operations define an approximately orthogonal factorization: s captures “what is where” while a captures “how it looks.” AdaIN then recombines these factors by rescaling normalized scene features with appearance-derived affine parameters (γ,β)(γ,β). This factorization is well-suited for registration tasks where domain shift is predominantly global (illumination, acquisition-specific intensity profiles) rather than spatially varying. The 32-dimensional appearance code provides sufficient capacity for such global shifts, while the 64×H×W64× H× W scene tensor preserves fine-grained spatial details. The strong quantitative results (FIRE: SSIM = 0.928, PSNR = 33.47 dB) validate this design, demonstrating that the factorization successfully isolates spatial structure from acquisition-specific appearance variations. 3.5 Cross-Domain Transfer: HPatches-Reg-256 A critical test of the factorization hypothesis is whether the same architecture generalizes to a fundamentally different imaging domain without architectural modification. We evaluate on HPatches-Reg-256 [3], which presents synthetic textured patches with Gaussian blob structures and structured noise, paired by random affine transformations (±15∘± 15 rotation, ±20± 20 px translation, 0.850.85–1.15×1.15× scaling, ±10∘± 10 shear). Unlike deformation-based methods [6, 13] that require domain-specific tuning of the deformation field parameterization, our factorization-based framework transfers directly. Table 3: Cross-domain applicability on HPatches-Reg-256 [3] (500 test pairs). All deep learning methods are trained from scratch on HPatches-Reg-256 for 30 epochs. Best in bold. Method NCC↑ SSIM↑ PSNR↑ Unregistered 0.312 0.241 14.87 VoxelMorph [2] 0.448 0.376 18.53 TransMorph [4] 0.471 0.395 19.24 SAS-Net [23] 0.502 0.421 20.15 GPEReg-Net (Ours) 0.536 0.450 21.01 GPEReg-Net achieves the highest scores across all three metrics on HPatches-Reg-256, with NCC of 0.536, SSIM of 0.450, and PSNR of 21.01 dB. Compared to the strongest baseline SAS-Net (PSNR = 20.15 dB), GPEReg-Net obtains a 0.86 dB improvement, confirming that the scene-appearance disentanglement framework transfers effectively across fundamentally different imaging domains. Deformation-based methods (VoxelMorph, TransMorph) achieve lower performance on this benchmark due to the large affine transformations (±15∘± 15 rotation, ±20± 20 px translation), which exceed the capacity of their deformation fields. fig. 2 presents qualitative registration outputs with SSIM annotations, and fig. 3 shows red-green overlay visualizations where reduced color separation indicates improved alignment. Figure 2: Qualitative registration results on HPatches-Reg-256. Four representative test samples showing fixed, moving, registered output, and absolute difference map with per-sample SSIM annotations. Figure 3: Red-green overlay comparison on HPatches-Reg-256 before and after registration (red = fixed, green = moving/registered). Reduced color separation indicates better spatial alignment. 4 Conclusion We have presented a principled factorization approach to cross-domain image registration: by decomposing images into domain-invariant scene representations and global appearance statistics, and recombining them via AdaIN, registration is achieved without estimating deformation fields. The position-encoded temporal attention module further enables the model to exploit inter-frame coherence in sequential acquisitions. Experiments on two diverse benchmarks—FIRE-Reg-256 (retinal fundus, semi-rigid) and HPatches-Reg-256 (synthetic textures, affine)—demonstrate that the factorization generalizes across imaging domains and deformation types, achieving state-of-the-art SSIM and PSNR while maintaining 1.87×1.87× faster inference than SAS-Net. Limitations and future work. The current appearance model ∈ℝ32 a ^32 captures only global intensity statistics; spatially-varying domain shifts (e.g., local illumination gradients) may require a spatially-conditioned appearance map () a( x). The fixed embedding table ∈ℝN×64 E ^N× 64 limits generalization to sequences longer than N frames; adaptive or continuous position encodings could address this. Integration with energy-based reconstruction frameworks [29] for learned feature compression is a promising direction. References [1] Avants, B.B., Epstein, C.L., Grossman, M., Gee, J.C.: Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain. Med. Image Anal. 12(1), 26–41 (Feb 2008) [2] Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J., Dalca, A.V.: VoxelMorph: A learning framework for deformable medical image registration. IEEE Trans. Med. Imaging 38(8), 1788–1800 (Aug 2019) [3] Balntas, V., Lenc, K., Vedaldi, A., Mikolajczyk, K.: HPatches: A benchmark and evaluation of handcrafted and learned local descriptors. In: Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR). p. 5173–5182 (2017) [4] Chen, J., Frey, E.C., He, Y., Segars, W.P., Li, Y., Du, Y.: TransMorph: Transformer for unsupervised medical image registration. Med. Image Anal. 82, 102615 (Nov 2022) [5] Chen, J., Liu, Y., Wei, S., Bian, Z., Subramanian, S., Carass, A., Prince, J.L., Du, Y.: A survey on deep learning in medical image registration: New technologies, uncertainty, evaluation metrics, and beyond. Med. Image Anal. 100, 103385 (Feb 2025) [6] de Vos, B.D., Berendsen, F.F., Viergever, M.A., Sokooti, H., Staring, M., Išgum, I.: A deep learning framework for unsupervised affine and deformable image registration. Med. Image Anal. 52, 128–143 (2019) [7] He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR). p. 770–778 (2016) [8] Hernandez-Matas, C., Zabulis, X., Triantafyllou, A., Anyfanti, P., Douma, S., Argyros, A.A.: FIRE: Fundus image registration dataset. In: Modelling the Physiological Human. p. 1–7. Springer (2017) [9] Horn, B.K., Schunck, B.G.: Determining optical flow. Artif. Intell. 17(1-3), 185–203 (Aug 1981) [10] Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proc. IEEE Int. Conf. Comput. Vis. (ICCV). p. 1501–1510 (2017) [11] Jaderberg, M., Simonyan, K., Zisserman, A., Kavukcuoglu, K.: Spatial transformer networks. In: Advances in Neural Information Processing Systems (NeurIPS). p. 2017–2025 (2015) [12] Lowe, D.G.: Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis. 60(2), 91–110 (Nov 2004) [13] Mok, T.C.W., Chung, A.C.S.: Large deformation diffeomorphic image registration with Laplacian pyramid networks. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2020. p. 211–221. Springer (2020) [14] Qin, J., Xu, Y., Lu, Z., Zhang, X.: Alternative telescopic displacement: An efficient multimodal alignment method. arXiv preprint arXiv:2306.16950 (2023) [15] Qin, J.: Bio-inspired Mamba: Temporal locality and bioplausible learning in selective state space models. arXiv preprint arXiv:2409.11263 (2024) [16] Qin, J.: Zoom and shift are all you need. arXiv preprint arXiv:2406.08866 (2024) [17] Qin, J., Liu, F., Zong, L.: Ancestral Mamba: Enhancing selective discriminant space model with online visual prototype learning for efficient and robust discriminant approach. arXiv preprint arXiv:2503.22729 (2025) [18] Qin, J., Liu, F., Zong, L.: BC-PMJRS: A brain computing-inspired predefined multimodal joint representation spaces for enhanced cross-modal learning. Neural Networks 188, 107449 (Apr 2025) [19] Qin, J., Peng, B., Liu, F., Cheng, G., Zong, L.: DUAL: Dynamic uncertainty-aware learning. arXiv preprint arXiv:2506.03158 (2025) [20] Qin, J., Liu, K., Cai, Y., Ji, T., Liu, F.: MTLP-MDG: Multi-task learning framework using probabilistic distribution perception for missing data generation. In: 2025 International Joint Conference on Neural Networks (IJCNN). p. 1–8 (2025) [21] Qin, J., Wang, Y.: Learning domain-invariant representations for cross-domain image registration via scene-appearance disentanglement. arXiv preprint arXiv:2601.08875 (2026) [22] Qin, J.: Progressive contrast registration for high-fidelity bidirectional photoacoustic microscopy alignment. arXiv preprint (2026) [23] Qin, J.: SAS-Net: Scene-appearance separation network for cross-domain spatiotemporal registration (2026), https://arxiv.org/abs/2602.09050 [24] Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2015. p. 234–241. Springer (2015) [25] Sotiras, A., Davatzikos, C., Paragios, N.: Deformable medical image registration: A survey. IEEE Trans. Med. Imaging 32(7), 1153–1190 (2013) [26] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems (NeurIPS). p. 5998–6008 (2017) [27] Vercauteren, T., Pennec, X., Perchant, A., Ayache, N.: Diffeomorphic demons: Efficient non-parametric image registration. NeuroImage 45(1), S61–S72 (Mar 2009) [28] Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE Trans. Image Process. 13(4), 600–612 (Apr 2004) [29] Wang, Y., Qin, J.: DCER: Dual-stage compression and energy-based reconstruction. arXiv preprint arXiv:2602.04904 (2026)