Frequency-Spatial Adaptive Selector (FSAS)

• FSAS is a family of adaptive selection mechanisms that couple frequency-sensitive representations with spatially conditioned modulation for various vision tasks.
• In image restoration tasks like demoiréing and deblurring, FSAS dynamically reweight critical spatial-frequency components, yielding measurable improvements in metrics such as PSNR.
• In stereo matching, FSAS integrates recurrent updates with contextual spatial attention to selectively fuse high- and low-frequency information, preserving edge details and overall structure.

Frequency-Spatial Adaptive Selector (FSAS) denotes a family of adaptive selection mechanisms that couple frequency-sensitive representations with spatially conditioned modulation. In recent arXiv literature, however, the acronym is not standardized. In "MoiréNet: A Compact Dual-Domain Network for Image Demoiréing," FSAS is a lightweight bottleneck selector that modulates dual-domain features before decoding (Guo et al., 23 Sep 2025). In "Efficient Frequency Domain-based Transformers for High-Quality Image Deblurring," FSAS instead denotes a frequency domain-based self-attention solver that replaces quadratic spatial-domain attention with FFT-based correlation (Kong et al., 2022). In "Selective-Stereo: Adaptive Frequency Information Selection for Stereo Matching," the closest FSAS formulation is the coupled use of a Selective Recurrent Unit and Contextual Spatial Attention to fuse high- and low-frequency hidden disparity updates (Wang et al., 2024). This suggests that FSAS is best understood as a recurring architectural pattern rather than a single canonical operator.

1. Terminological scope and recurrent design pattern

Across the cited literature, FSAS consistently refers to adaptive selection over information that differs in frequency content and spatial relevance, but the exact computational form varies by task and backbone. In restoration, FSAS may be a bottleneck refinement block or an FFT-based attention surrogate; in stereo matching, it may be a selector over recurrent updates; in related work, the same idea appears under different names such as frequency selection, adaptive fusion, or waveform switching.

Work	Meaning of FSAS or closest equivalent	Primary role
MoiréNet	Frequency-Spatial Adaptive Selector	Bottleneck feature selection and reweighting
FFTformer	frequency domain-based self-attention solver	Efficient attention via frequency-domain correlation
Selective-Stereo	SRU + CSA as an FSAS-style selector	Pixel-adaptive fusion of multi-frequency recurrent updates

A common misconception is that FSAS always denotes a self-attention block. The literature does not support that reading. In MoiréNet, FSAS is explicitly a lightweight selection-and-reweighting block positioned after feature extraction and before decoding; in FFTformer, FSAS is the attention solver itself; and in Selective-Stereo, the selector is implemented as spatial attention over recurrent branches rather than as token attention (Guo et al., 23 Sep 2025, Kong et al., 2022, Wang et al., 2024).

2. MoiréNet FSAS as bottleneck dual-domain feature selection

In MoiréNet, FSAS is the paper’s lightweight selection-and-reweighting block for deciding which dual-domain features should be emphasized at the bottleneck of a dual-domain U-Net-like demoiréing network. Its role is explicitly downstream of feature extraction: the encoder side first uses DFSE to model the directional and periodic structure of moiré patterns, and FSE to model general spatial-frequency content, after which FSAS acts as a feature selector/refiner before decoding (Guo et al., 23 Sep 2025).

The module is described as enhancing “critical spatial-frequency components” and “dynamically emphasize important regions.” Its operation is sequential. Given an input feature map $F$, FSAS first derives a spatial importance map $F_s$ using max pooling and average pooling along the channel dimension, followed by depth-wise convolution. It then applies a mean-centering step and residual spatial modulation: $$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$ This makes the selector residual and mean-centered: the original $F$ is added back, and the spatial gate emphasizes deviations from the average response rather than uniformly scaling the feature map.

After spatial selection, FSAS performs pixel-level refinement. The paper states that it concatenates $F_{fs}$ with the pixel-attention response, applies grouped convolutions, and generates pixel-wise weights $W_p$ using a sigmoid activation. The final output is formed by residual modulation with learnable scalars: $$F_{\text{out}} = a\big(W_p \otimes F_{fs}\big) + bF ,$$ where $a$ and $b$ are initialized to $0$ and $F_s$0, respectively. The intended meaning is that the module starts as an identity-like mapping and learns to gradually introduce adaptive reweighting, which the paper associates with training stability and resistance to early over-modulation.

The distinction between DFSE and FSAS is architecturally central. DFSE is the earlier structural encoder, containing FSE, DAC, and DRU, and is designed to explicitly model moiré orientation and periodicity. FSAS, by contrast, does not attempt to build the main representation from scratch; it takes already-extracted features and decides what to emphasize. The division of labor is therefore: DFSE determines what features to extract, and FSAS determines which extracted features matter most for restoration.

The ablation evidence on the FHDMi dataset quantifies this contribution. The baseline without FSAS or FAM is 21.79 dB PSNR / 0.7922 SSIM. Adding only FSAS gives 22.26 dB / 0.7981; adding only FAM gives 22.30 dB / 0.8005; and using both together gives 22.51 dB / 0.8062. This establishes a gain of about +0.47 dB PSNR for FSAS over the baseline. The total MoiréNet size is 5.513M parameters, but the paper does not report a separate parameter breakdown for FSAS alone (Guo et al., 23 Sep 2025).

3. FFTformer FSAS as a frequency-domain self-attention solver

In FFTformer, FSAS has a different expansion and a different computational objective. It is a frequency domain-based self-attention solver developed for image deblurring, motivated by the convolution theorem: correlation or convolution in the spatial domain is equivalent to an element-wise product in the frequency domain. The paper treats scaled dot-product attention as fundamentally a correlation between query and key tokens, and uses that observation to estimate attention in the frequency domain rather than by explicitly forming the quadratic attention matrix (Kong et al., 2022).

The standard scaled dot-product attention forms $F_s$1, whose space and time complexity are reported as $F_s$2 and $F_s$3. FSAS avoids this bottleneck. It first obtains $F_s$4, $F_s$5, and $F_s$6 via a $F_s$7 point-wise convolution and a $F_s$8 depth-wise convolution. It then applies FFT to $F_s$9 and $$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$0, estimates their correlation in the frequency domain by FFT, conjugation, multiplication, and inverse FFT, denotes the resulting attention-like map by $$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$1, and computes

$$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$2

followed by the residual output

$$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$3

The complexity claim is one of the module’s defining properties. Compared with standard spatial-domain self-attention, FSAS is reported to reduce the cost to $$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$4 space complexity and $$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$5 time complexity; the introduction also states that the space and time complexity can be reduced to $$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$6 and $$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$7 for each feature channel, depending on how the cost is counted. Architecturally, the module is used only in the decoder of an asymmetric encoder-decoder network. The reason is explicit: shallow encoder features still contain strong blur, so similarity estimation on those features may be unreliable, whereas deep decoder features are clearer and more semantically reliable.

FFTformer couples FSAS with DFFN, a discriminative frequency domain-based feed-forward network. In this pairing, FSAS provides efficient global correlation modeling and DFFN selectively preserves useful low- and high-frequency content. The ablations support both the efficiency and placement claims. The memory/runtime table reports about 42–44 ms on a $$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$8 image with about 5.9–6.5G GPU memory across several window sizes, while the window-based baseline becomes slower and eventually runs out of memory for larger windows. In restoration quality, “FSAS+FFN” exceeds “w/ only FFN” by 0.42 dB PSNR. “SA w/ SD” is reported to be 0.27 dB lower than the full FSAS+DFFN model. The location ablation further shows that using FSAS in both encoder and decoder yields 33.56 dB / 0.9653 SSIM, while using FSAS only in the decoder improves this to 33.73 dB / 0.9663 SSIM (Kong et al., 2022).

4. Selective-Stereo and FSAS as adaptive recurrent frequency fusion

In stereo matching, the FSAS concept appears as adaptive frequency information selection within an iterative update operator. The proposed mechanism, referred to in the synthesis as FSAS = SRU + CSA, replaces the fixed-receptive-field recurrent update used in methods such as RAFT-Stereo and IGEV-Stereo. The core claim is that edge regions and smooth or ambiguous regions require different frequency information: high-frequency information is needed in edges, thin objects, and detailed structures, whereas low-frequency information is needed in textureless, smooth, ambiguous, or reflective regions (Wang et al., 2024).

The Selective Recurrent Unit (SRU) produces candidate hidden updates using multiple GRU branches with different kernel sizes, which correspond to different receptive fields and therefore different frequency characteristics. The Contextual Spatial Attention (CSA) module generates a spatially varying attention map from context information. The fusion rule is

$$F_{fs} = (F \otimes (F_s - \mathrm{mean}(F_s))) + F .$$9

where $F$0 is the attention map from CSA, $F$1 is the GRU output with smaller kernels, and $F$2 is the GRU output with larger kernels. This is the central selector equation: the small-kernel branch favors high-frequency detail, the large-kernel branch favors low-frequency or contextual information, and CSA decides their relative importance pixel-wise.

CSA itself contains Channel Attention Enhancement and Spatial Attention Extractor modules, simplified from CBAM. It uses average pooling and max pooling over spatial dimensions to generate channel weights, then pools along the channel dimension, concatenates the pooled maps into $F$3, and applies one convolution plus sigmoid to produce the final attention map. The paper states the semantics of this map directly: regions needing high-frequency information get high weights, and regions needing low-frequency information get low weights.

The empirical evidence indicates that the selector is not merely adding more branches; it is using them correctly. On Scene Flow, the ablation table reports the following for RAFT-Stereo: baseline EPE 0.53, >1px 6.08, 11.12M parameters; SRU only EPE 0.50, >1px 5.38, 11.65M; SRU + CSA (Contrary) EPE 0.50, >1px 5.58, 11.65M; and full Selective-RAFT EPE 0.47, >1px 5.32, 11.65M. The “Contrary” setting is important because reversing the attention harms performance relative to normal CSA, which supports the claim that the learned attention map is aligned with regional frequency needs. Region-wise evaluation further reports improvement in edge regions from 3.21 to 2.40 EPE and in non-edge regions from 0.53 to 0.40. The kernel size ablation identifies $F$4 as the best default combination (Wang et al., 2024).

5. Related selector mechanisms not literally named FSAS

Several recent systems implement closely related ideas without using the exact FSAS label. In MSFSNet, the relevant mechanisms are DFS, FCAM, and SFF. DFS dynamically generates low- and high-frequency maps with a learnable filter, FCAM uses bidirectional cross-attention to determine which low-frequency and high-frequency information should be retained, and SFF selectively identifies the information to be transmitted in the skip connection, mitigating the interference of implicit noise. The paper describes this family of modules as “selectively recovering richer and more accurate information” and “adaptively selects the most informative frequency for recovery.” In ablations, the baseline is 32.83 dB, adding multi-input/multi-output gives +0.09 dB, adding SFF gives +0.21 dB, adding DFS gives +0.33 dB, and all modules together yield 33.42 dB, a +0.59 dB gain. Plug-and-play insertion into NAFNet gives +0.14 dB for SFF, +0.30 dB for DFS, and +0.44 dB for both (Gao et al., 2024).

A different but conceptually parallel use of adaptive selection appears in wireless communications. The paper on real-time parameter estimation for adaptive OFDM/OTFS selection does not use the term FSAS, but it proposes a real-time parameter-estimation-driven waveform switcher based on an ISAC radar front-end. Delay, Doppler, speed, and range are estimated, converted by

$F$5

and compared against the threshold

$F$6

The decision rule is

$F$7

The selector is therefore a mobility-adaptive waveform controller rather than a learned feature module. OFDM complexity is reported as $F$8, OTFS complexity as $F$9, and the MF-F estimator complexity as approximately $F_{fs}$0. The qualitative conclusion is explicit: OFDM wins in low mobility, OTFS wins in high mobility, and the hybrid selector wins overall (Darghouthi et al., 2024).

In small object detection, FSDC-DETR introduces DBFSAF, identified in the synthesis as the closest equivalent to an FSAS module. DBFSAF concatenates dual-branch features, applies FDConv, splits channels with a partial ratio $F_{fs}$1, refines the selected channels with MKSRU and FPU, derives adaptive weights by GAP and Softmax, and fuses the result with a preserved path. The best partial ratio is $F_{fs}$2. At the system level, the paper reports that FSDC-DETR improves AP by 6.4 on VisDrone-DET2019 and 6.6 on AITODv2, with gains of 6.8 and 6.9 AP for small objects (Liu et al., 6 Jul 2026).

6. Comparative interpretation, misconceptions, and significance

The cross-paper record supports three clarifications. First, FSAS is not a standardized layer. The same acronym names materially different mechanisms in MoiréNet and FFTformer, and in Selective-Stereo the term is most accurate as a conceptual shorthand for SRU plus CSA rather than as the paper’s own principal module name (Guo et al., 23 Sep 2025, Kong et al., 2022, Wang et al., 2024).

Second, FSAS is not reducible to “frequency processing.” Every instance couples frequency-sensitive behavior with spatial, contextual, or pixel-level conditioning. MoiréNet uses channel pooling, depth-wise convolution, grouped convolution, sigmoid gating, residual addition, and learnable scalar mixing. FFTformer uses FFT-based correlation but still returns to spatially structured feature aggregation through normalization, value projection, and residual $F_{fs}$3 convolution. Selective-Stereo uses context-derived spatial attention to choose between small- and large-kernel recurrent updates. Related systems such as MSFSNet and FSDC-DETR likewise separate, refine, weight, and then merge frequency and spatial pathways (Gao et al., 2024, Liu et al., 6 Jul 2026).

Third, the functional objective of FSAS varies with the failure mode of the task. In demoiréing, the selector is designed to adaptively suppress moiré-corrupted responses while preserving structural detail. In deblurring, it is designed to compute global correlations efficiently without the quadratic bottleneck of standard attention. In stereo matching, it is designed to preserve edge detail without losing the broader context required in low-texture regions. In communications, the analogous selector is not a feature gate at all, but a threshold-based controller that switches between OFDM and OTFS according to sensed mobility. This suggests that “frequency-spatial adaptive selection” is a transferable principle whose concrete implementation is highly task-specific.

From an architectural perspective, the shared significance of FSAS-like mechanisms lies in selective modulation rather than universal structure. The literature repeatedly emphasizes adaptivity, residual preservation, lightweight or low-complexity design, and the need to avoid uniformly treating all locations, all frequency bands, or all operating conditions as equivalent. In that sense, FSAS is less a fixed module family than an organizing idea for designing selectors that decide what to keep, what to attenuate, and where that decision should depend on frequency content, spatial context, or both.