SRSNet: Segmentation & Forecasting Models

• SRSNet is a shared acronym for two distinct architectures: one for weak-target image segmentation using a Siamese reconstruction branch with DPConv, and another for long-term forecasting using adaptive patch selection.
• The segmentation model employs dual-branch encoder-decoders where dynamic-parameter convolution transfers reconstruction features to enhance accuracy on medical and infrared datasets.
• The forecasting model utilizes a Selective Representation Space to adaptively select and reorder patch tokens, yielding competitive results across multivariate time series benchmarks.

Searching arXiv for papers using the name “SRSNet” to verify naming and disambiguation. SRSNet is a model name that appears in distinct arXiv lines of work rather than denoting a single canonical architecture. In weak-target image segmentation, it refers to the Siamese Reconstruction-Segmentation Network, a dual-branch encoder-decoder architecture coupled through Dynamic-Parameter Convolution (DPConv) and introduced for medical and infrared image segmentation (Nian et al., 2023). In long-term time series forecasting, it refers to a forecasting model built from a Selective Representation Space (SRS) module and a lightweight Linear/MLP head, with the central objective of adaptively selecting and reordering patch tokens before prediction (Wu et al., 16 Oct 2025). A recurrent source of confusion is the similarly named SSRNet for hyperspectral image super-resolution, which is explicitly not called SRSNet in its source paper (Wang et al., 2020).

1. Nomenclature and scope

The term SRSNet is used for at least two technically unrelated architectures. The segmentation paper explicitly introduces SRSNet, short for Siamese Reconstruction-Segmentation Network, and the title variant SRSNetwork expands this as “Siamese Reconstruction-Segmentation Networks based on Dynamic-Parameter Convolution” (Nian et al., 2023). The forecasting paper defines SRSNet as a simple forecasting architecture consisting of the Selective Representation Space (SRS) module plus a Linear/MLP prediction head (Wu et al., 16 Oct 2025).

A concise disambiguation is useful because the shared acronym conceals different task formulations, data modalities, and optimization procedures.

Usage of “SRSNet”	Domain	Core mechanism
Siamese Reconstruction-Segmentation Network	Weak target image segmentation	Reconstruction branch generates dynamic kernels for segmentation via DPConv
SRSNet built on Selective Representation Space	Long-term time series forecasting	Adaptive patch selection, dynamic reordering, and embedding fusion
SSRNet	Hyperspectral image super-resolution	Different model name; not SRSNet

This naming overlap suggests that “SRSNet” should be interpreted contextually from the problem domain. In computer vision segmentation, it denotes a reconstruction-guided segmentor; in forecasting, it denotes a patch-space constructor coupled to a shallow predictor. The hyperspectral SR paper is relevant mainly because it explicitly states that SSRNet—not SRSNet—is the model name, and any use of “SRSNet” there would be mistaken (Wang et al., 2020).

2. SRSNet as a Siamese Reconstruction-Segmentation Network

In weak target image segmentation, SRSNet is designed for settings in which targets exhibit weak internal structure, unclear boundaries, low contrast, or strong background clutter. The paper emphasizes medical image segmentation, especially ultrasound and CT, and infrared small/weak target segmentation as representative cases (Nian et al., 2023). Its motivating claim is that reconstruction and segmentation should be treated as dual tasks: reconstruction is a low-level task that can absorb domain-relevant image statistics from unlabeled data, while segmentation is a high-level task that can benefit from those learned priors.

Architecturally, SRSNet is a dual-branch encoder-decoder system with an upper reconstruction branch, a lower segmentation branch, and DPConv as the coupling mechanism. The architecture is described as enhanced from DNANet, with ResBlocks, CBAM attention, and BLAM from ALCNet after the encoder-decoder stage (Nian et al., 2023). The paper also states that the branches are not strictly identical: the reconstruction network uses ResBlock counts $[5,5,5,5]$, whereas the segmentation network uses $[6,8,12,6]$. Accordingly, the label “siamese” is nonclassical here. It denotes a paired and coupled two-branch design rather than the usual strict notion of two identical subnetworks with shared weights.

The central coupling is that the reconstruction branch provides features $Y$ used to generate dynamic convolution kernels, while the segmentation branch provides features $X$ that are convolved by those generated kernels. The paper does not state that ordinary convolution weights are shared between the branches. It also does not provide a fully specified per-layer placement schedule for DPConv, though the figures and text indicate that DPConv connects encoder and decoder features across the two tasks.

The stated contributions are threefold: DPConv as a dynamic-parameter convolution operator adaptive to the input-feature distribution; SRSNet as a siamese reconstruction-segmentation architecture that can improve segmentation by increasing reconstruction-task training data rather than segmentation labels; and extensive experiments on seven datasets, including five medical and two infrared datasets, against 16 baselines (Nian et al., 2023).

3. Dynamic-Parameter Convolution and the reconstruction–segmentation coupling

The key low-level operator in the segmentation SRSNet is Dynamic-Parameter Convolution (DPConv). The paper positions DPConv against prior dynamic convolution variants such as CondConv, DyConv, and ODConv, which it summarizes as learning a weighted combination of a fixed kernel bank: $$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$ In that characterization, the effective kernel remains inside the span of pretrained kernels $\{\omega_i\}$, with input-conditioned coefficients $\alpha_{ji}$. The authors argue that such mechanisms are not “truly dynamic” because they recombine static filters rather than synthesize each parameter directly.

DPConv instead generates the convolution kernel parameters from features. Its defining equation is

$$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$

where $X$ denotes segmentation-task information, $Y$ denotes reconstruction-task information, $[6,8,12,6]$0 transforms the segmentation features, $[6,8,12,6]$1 applies convolution, average pooling, and convolution to reconstruction features, and $[6,8,12,6]$2 maps reconstruction features to the actual kernel parameters (Nian et al., 2023). In operational terms, the reconstruction branch predicts the convolution weights and those weights are applied to segmentation features.

For a convolution with $[6,8,12,6]$3 input channels, $[6,8,12,6]$4 output channels, and kernel size $[6,8,12,6]$5, the total number of generated parameters is

$[6,8,12,6]$6

The kernel is represented as $[6,8,12,6]$7. To obtain $[6,8,12,6]$8 from reconstruction features $[6,8,12,6]$9, the paper gives

$Y$0

Here, $Y$1 denotes a $Y$2 convolution from $Y$3 to $Y$4 channels and $Y$5 denotes adaptive average pooling. The paper reports that a direct average-based mapping from $Y$6 to $Y$7 caused non-convergence, motivating the two-step $Y$8-conv + pooling + $Y$9-conv parameter generator.

This construction encodes the paper’s broader thesis about task duality. Reconstruction learns image distribution and semantic-statistical structure, while segmentation requires precisely such information in weak-target regimes. DPConv serves as the transfer operator that turns reconstruction features into adaptive priors usable by the segmentor. A plausible implication is that the paper treats the reconstruction branch not as a standard auxiliary head, but as a source of input-conditioned operators for the downstream segmentation process.

4. Optimization, datasets, and empirical profile of the segmentation model

The segmentation SRSNet is trained in two stages rather than by simultaneous end-to-end multi-task optimization. First, the reconstruction network is trained on image reconstruction in an unsupervised manner. Second, the segmentation network is trained with segmentation labels while the reconstruction branch remains in the forward path but is frozen and does not participate in backpropagation (Nian et al., 2023). The paper therefore does not provide a single joint objective of the form

$X$0

The reconstruction loss is given as $X$1 loss: $X$2 The segmentation loss is

$X$3

The implementation details reported are: SGD, weight decay $X$4, momentum $X$5, initial learning rate $X$6, poly learning-rate schedule, batch size $X$7, $X$8 epochs, and a single NVIDIA GeForce RTX4090 GPU. All images are resized to $X$9, with random rotation and flip as augmentation (Nian et al., 2023).

The seven datasets are BUS, BUSI, TNSCUI, ISIC2018, Synapse, SIRST, and NUDT-KBT19. For BUS, BUSI, TNSCUI, ISIC2018, SIRST, and NUDT-KBT19, the split is a random 70/30 train/validation split repeated three times; for Synapse, the setting follows TransUnet. Metrics are IoU and F1 for BUS, BUSI, TNSCUI, ISIC2018, SIRST, and NUDT-KBT19, and mIoU, Dice, and HD95 for Synapse (Nian et al., 2023).

The reported results are strongest on difficult weak-target settings. On BUS, SRSNet achieves 88.29 IoU / 93.44 F1; on BUSI, 74.47 / 82.57; on TNSCUI, 79.33 / 87.02; and on ISIC2018, 83.16 / 89.72, where the paper notes competitive rather than unambiguously dominant performance. On Synapse, SRSNet reports 69.28 mIoU, 79.80 Dice, and 24.51 HD95, exceeding TransUnet on mIoU and Dice while remaining second-best on HD95. On SIRST, it reaches 74.67 IoU / 84.28 F1, and on NUDT-KBT19, 86.34 / 92.62 (Nian et al., 2023).

The ablations attribute the main technical improvement to DPConv. In comparisons among S+DyConv, S+ODConv, S+ScConv, and SRS(DPConv), DPConv is consistently favored; on BUSI, for example, ScConv yields 71.29 IoU / 79.85 F1 while DPConv yields 74.47 / 82.57. A second ablation compares S Stage against R+S Stage, with moderate but positive gains such as 73.85 $$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$0 74.47 IoU on BUSI and 73.78 $$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$1 74.67 IoU on SIRST. The reconstruction-data ablation, comparing T dataset only against T + V dataset, is mixed; the authors explicitly note that the trend is “not always positively impactful” and seems more beneficial on larger datasets (Nian et al., 2023).

These results support the method’s intended use when segmentation labels are scarce and unlabeled images are easy to obtain. At the same time, the paper leaves several caveats explicit or implicit: reconstruction-data scaling is not uniformly beneficial, training is staged rather than jointly optimized, architectural details are partly underspecified, and efficiency claims are qualitative because no exact parameter-count, FLOP, or timing tables are reported (Nian et al., 2023).

5. SRSNet as a Selective Representation Space forecaster

In long-term time series forecasting, SRSNet has a different meaning. The forecasting paper studies supervised prediction from a contextual multivariate time series

$$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$2

to future values

$$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$3

where $$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$4 is the number of variables, $$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$5 is the look-back length, and $$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$6 is the prediction horizon (Wu et al., 16 Oct 2025). The method is explicitly motivated by patch-based forecasting, in which the historical input is partitioned into short subsequences whose embeddings serve as tokens. The paper’s critique is that conventional adjacent patching creates a fixed representation space because patches always come from predetermined positions.

The standard adjacent patching formulation is

$$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$7

with patch size $$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$8, stride $$\omega = \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_{ji}\omega_i.$$9, and $\{\omega_i\}$0 adjacent patches. SRS instead constructs a candidate stride-1 patch set

$\{\omega_i\}$1

and adaptively selects $\{\omega_i\}$2 patches from those $\{\omega_i\}$3 candidates (Wu et al., 16 Oct 2025). The authors emphasize the expressivity of this construction by noting a search space of

$\{\omega_i\}$4

possible patch multisets under sampling with replacement, and

$\{\omega_i\}$5

possible jointly selected and ordered patch sequences when Dynamic Reassembly is included.

Formally, SRSNet in this paper is SRS + a simple MLP head. The SRS module contains two key mechanisms: Selective Patching, which chooses informative patches from all stride-1 candidates, and Dynamic Reassembly, which adaptively orders those selected patches. The end-to-end pipeline is: Instance Normalization; standard adjacent patch construction; Selective Patching over all candidate stride-1 patches; Dynamic Reassembly; separate embedding of the conventional and selective patches; Adaptive Fusion of the two embedding streams; and prediction through a Linear/MLP head (Wu et al., 16 Oct 2025).

Selective Patching uses an MLP scorer

$\{\omega_i\}$6

with

$\{\omega_i\}$7

The design outputs $\{\omega_i\}$8 scores per candidate patch, corresponding to $\{\omega_i\}$9 sampling rounds, which permits selection with replacement. Because $\alpha_{ji}$0 is non-differentiable, the paper introduces a hard-but-gradient-carrying construction: $\alpha_{ji}$1

$\alpha_{ji}$2

$\alpha_{ji}$3

The forward pass is therefore hard and deterministic, while gradients flow through the attached score terms.

Dynamic Reassembly applies the same idea to ordering. Using

$\alpha_{ji}$4

the method computes

$\alpha_{ji}$5

followed by a parallel gradient-attachment construction to obtain reordered patches $\alpha_{ji}$6 (Wu et al., 16 Oct 2025). The paper’s stated intuition is that many patch-based backbones are permutation-variant, so useful but non-adjacent temporal regions may be better modeled after adaptive shuffling rather than strict chronological preservation.

The selective and conventional patch streams are then embedded: $\alpha_{ji}$7 and combined via

$\alpha_{ji}$8

After positional information is added,

$\alpha_{ji}$9

the SRSNet predictor flattens the representation and applies an MLP: $$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$0 with training loss

$$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$1

The head is reported as a Linear/MLP head ($$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$2 layers) (Wu et al., 16 Oct 2025).

6. Forecasting results, plugin behavior, and limitations

The forecasting SRSNet is evaluated on eight multivariate long-term forecasting datasets: ETTh1, ETTh2, ETTm1, ETTm2, Weather, Electricity, Solar, and Traffic (Wu et al., 16 Oct 2025). The comparison includes TimeKAN, Amplifier, iTransformer, TimeMixer, PatchTST, Crossformer, TimesNet, DLinear, Non-stationary Transformer, and FEDformer. The protocol reports MSE and MAE across forecasting horizons $$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$3, uses a unified TFB benchmark setup, and states that it does not apply the “Drop Last” trick.

The average results reported over the four horizons are: ETTh1 0.404 / 0.424, ETTh2 0.334 / 0.385, ETTm1 0.351 / 0.378, ETTm2 0.252 / 0.314, Weather 0.226 / 0.266, Electricity 0.161 / 0.254, Solar 0.183 / 0.239, and Traffic 0.392 / 0.270 (Wu et al., 16 Oct 2025). In the full horizon-by-horizon table, SRSNet records 23 firsts in MSE and 20 firsts in MAE. The paper interprets this as evidence that improving the representation space can outperform the strategy of placing a more complex backbone on top of a fixed adjacent patch partition.

A major emphasis of the paper is that SRS is a plug-and-play, backbone-agnostic module. It is evaluated as a plugin for MLP, PatchTST, Crossformer, xPatch, and PatchMLP (Wu et al., 16 Oct 2025). The reported average MSE improvements in the summary table are: for MLP $$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$4 SRSNet, +6.05\% on ETTh1, +7.70\% on ETTm2, +16.08\% on Solar, and +5.26\% on Traffic; for PatchTST + SRS, +3.48\%, +3.00\%, +8.87\%, and +2.74\%; for Crossformer + SRS, +1.55\%, +9.83\%, +5.66\%, and +1.85\%; for PatchMLP + SRS, +2.99\%, +2.94\%, +7.00\%, and +2.57\%; and for xPatch + SRS, +2.38\%, +3.59\%, +4.01\%, and +2.17\%.

The component ablations compare w/o SRS, w/o Selective Patching, w/o Dynamic Reassembly, w/o Adaptive Fusion, and full SRSNet, with the reported conclusion that Selective Patching has the greatest impact, while Dynamic Reassembly and Adaptive Fusion also contribute consistently (Wu et al., 16 Oct 2025). The decomposition is conceptually clear: selection determines what is represented, reassembly determines in what order it is represented, and fusion determines how much conventional adjacent-patch information is retained.

The training details reported include Adam, PyTorch, Python 3.8, NVIDIA Tesla-A800 GPU, Instance Normalization at input, initial batch size 64 with halving down to 8 if out-of-memory occurs, and look-back window sizes searched in $$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$5. Common hyperparameters in the sensitivity analysis are patch sizes 16 or 24, and 2 hidden layers with hidden dimension 128 for both scorers (Wu et al., 16 Oct 2025). The paper does not specify the learning rate, weight decay, number of epochs, or learning-rate schedule in the provided text.

Efficiency is treated explicitly. With look-back $$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$6, horizon $$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$7, and batch size $$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$8, on ETTh1 SRSNet reports 1012 MB and 2.27 s/batch, compared with 1404 MB and 2.49 s for PatchTST, 3976 MB and 17.13 s for Crossformer, 8190 MB and 64.83 s for FEDformer, and 2846 MB and 14.95 s for TimesNet. On Solar, SRSNet reports 6301 MB and 56.149 s/batch, compared with 26777 MB and 137.60 s for PatchTST, 16375 MB and 205.60 s for Crossformer, and 13109 MB and 326.38 s for TimeKAN (Wu et al., 16 Oct 2025). As a plugin, SRS is reported to add only modest overhead; for example, on ETTh1, PatchTST + SRS incurs memory +2.47\%, inference +12.73\%, training +12.31\%, and MACs +4.26\%, while Crossformer + SRS incurs memory +3.69\%, inference +10.21\%, training +8.17\%, and MACs +0.61\%.

The forecasting paper also states several limitations. SRS is not practical for non-patch models; its behavior under scaling law / foundation model settings is unverified; the selected patches are not guaranteed to be human-interpretable; and the initialization of the fusion weight $$DPConv(X,Y) = Conv2d(\psi_{1}(X),\psi_{2}(\gamma(Y))),$$9 appears important, with larger $X$0 recommended for more periodic or stationary datasets and smaller $X$1 for more non-stationary or shifting data when prior knowledge is available (Wu et al., 16 Oct 2025).

7. Comparative interpretation

Taken together, the two uses of SRSNet illustrate a shared naming pattern but not a shared technical lineage. In the segmentation paper, the acronym encodes a siamese reconstruction-segmentation design in which unsupervised reconstruction informs supervised segmentation by generating dynamic kernels (Nian et al., 2023). In the forecasting paper, the acronym encodes a model built around Selective Representation Space, whose contribution is to replace fixed adjacent patching with adaptive patch selection, ordering, and fusion (Wu et al., 16 Oct 2025).

The conceptual commonality is that both models intervene on the representation interface rather than merely scaling a standard backbone. The segmentation SRSNet changes how convolutional operators are parameterized, using reconstruction features as the source of dynamic kernels. The forecasting SRSNet changes how input tokens are constructed, using hard selection and hard sorting over candidate patches before any downstream prediction. This suggests a broader editor’s term, “representation-space intervention,” for the family resemblance between the two usages: both methods seek gains by modifying the information representation made available to a downstream predictor. That phrase is interpretive rather than part of either source paper.

The principal misconception to avoid is conflating either SRSNet with SSRNet for hyperspectral image super-resolution. The hyperspectral SR paper repeatedly names its method SSRNet, not SRSNet, and attributes its performance to 3D spatial-spectral modeling, the SSRM module, local feature fusion, and residual learning rather than to either DPConv or selective patch reassembly (Wang et al., 2020). In bibliographic or technical discussion, explicit expansion of the acronym is therefore essential.