Spatial Enhanced Modules in Deep Learning

• Spatial Enhanced Module is a neural or signal-processing component that refines local details by reweighting spatial features, injecting geometric priors, and fusing complementary cues.
• It employs methods like convolutional attention, transformer-based Euclidean decay, and group-wise semantic sharpening to enhance feature extraction and boundary precision.
• When integrated into CNNs, Transformers, and hybrid architectures, these modules offer improved spatial sensitivity and efficiency with minimal computational overhead.

Searching arXiv for papers using or defining “Spatial Enhanced Module” and closely related spatial enhancement modules. I’ll query arXiv for the exact phrase and adjacent module names to ground the article in the primary literature. A spatial enhanced module is a neural or signal-processing component that strengthens the modeling of spatial structure by reweighting locations, injecting explicit geometric priors, enlarging or reshaping receptive fields, or fusing spatial cues with complementary domains such as frequency, time, channel, or text. Across recent literature, the expression does not denote a single canonical operator. Instead, it refers to a design category spanning convolutional attention blocks, transformer attention with distance decay, group-wise semantic enhancement, masked region-specific harmonization, frequency-aware spike refinement, selective scan modules, and reconstruction frameworks that use an auxiliary grayscale path to recover fine spatial detail. This suggests that the term functions less as a fixed architecture than as a recurring response to a common problem: baseline backbones often model global context or semantic abstraction effectively, yet remain limited in local detail recovery, boundary precision, spatial bias, or geometry awareness (Wang et al., 2023, Woo et al., 2018, Li et al., 2019, Fan et al., 20 Apr 2026, Sabharwal et al., 2024, Huang et al., 24 Jun 2025).

1. Terminological scope and problem setting

In computer vision and related fields, spatial enhancement is typically introduced when the base architecture is judged insufficiently sensitive to where salient information lies. In CBAM, the spatial branch is motivated by the observation that convolutional features encode both “what” is important and “where” it is important; the module therefore learns a 2D mask over spatial positions after channel refinement (Woo et al., 2018). In SGE, the difficulty is different: semantic sub-features are often spatially affected by similar patterns and noisy backgrounds, so the module generates an attention factor for each spatial location in each semantic group, allowing each group to autonomously enhance its learnt expression and suppress possible noise (Li et al., 2019). In image harmonization, S$^2$AM is based on the empirical observation that the non-spliced region should stay essentially unchanged and the main difference lies in the spliced region, so the network should separate the foreground/spliced area from the background and learn their feature transformations differently (Cun et al., 2019).

Transformer-oriented work uses the same broad label for a different deficiency. EVT argues that self-attention lacks explicit spatial priors and therefore augments it with Euclidean-distance decay and a spatially-independent grouping strategy so that attention becomes geometry-aware while remaining flexible in token grouping (Fan et al., 20 Apr 2026). SCSC similarly frames its contribution as a general module for both CNNs and Transformers, motivated by the claim that large dense kernels and self-attention are inefficient and that large receptive fields make it hard to capture local features; its solution is an efficient spatial cross-scale encoder and spatial embed module that capture assorted features in one layer (Wang et al., 2023).

The same design pressure appears outside standard image classification. TexLiverNet introduces a Spatial-Frequency Perception Module because repeated downsampling causes loss of fine image features, DOR-MLP “lacks precision in capturing local details,” and skip connections may introduce noise (Jiang et al., 2024). FSTA-SNN is motivated by depth-wise changes in spatial frequency preference in spiking networks and uses a DCT-based spatial branch to suppress redundant spike features (Yu et al., 2024). In remote sensing change detection, 3D-SSM extends selective scanning to the spatial plane and channel perspectives because existing Mamba-based methods do not sufficiently capture long-range dependencies between image channels (Huang et al., 24 Jun 2025). The term therefore spans a family of remedies for inadequate spatial inductive bias, inadequate spatial selectivity, or inadequate spatial-frequency coupling.

2. Canonical design principles

One canonical pattern is sequential channel-then-spatial refinement. CBAM first computes channel attention and then spatial attention: $$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$ Its spatial attention compresses the channel dimension with average pooling and max pooling, concatenates the resulting descriptors, applies a $7\times 7$ convolution, and then a sigmoid: $$\mathbf{M_s}(\mathbf{F})=\sigma\left(f^{7\times 7}\left([AvgPool(\mathbf{F}); MaxPool(\mathbf{F})]\right)\right).$$ The principle is economical: channel compression precedes the spatial operator, so spatial context is enlarged without a deep or expensive subnetwork (Woo et al., 2018).

A second pattern is group-wise semantic sharpening. SGE divides channels into groups, computes a global descriptor for each group, measures the similarity between the group descriptor and every local spatial vector, normalizes the similarity map, and gates the original features: $$g=\frac{1}{m}\sum_{i=1}^{m}x_i, \qquad c_i=g\cdot x_i, \qquad \hat{x}_i=x_i\cdot \sigma(a_i).$$ Because the similarity is defined relative to a group-level semantic prototype, each semantic group can highlight different regions. The mechanism is lightweight because the only learnable parameters are the affine coefficients $\gamma$ and $\beta$ per group (Li et al., 2019).

A third pattern is explicit spatial prior injection rather than attention-map learning alone. EVT modifies self-attention through a Euclidean-distance decay matrix: $$\mathrm{EuSA}(X) = (\mathrm{Softmax}(QK^T)\odot E^{2d})V,\qquad E_{nm}^{2d}=\gamma^{\sqrt{(x_n-x_m)^2+(y_n-y_m)^2}}.$$ The attention contribution from distant tokens is therefore attenuated directly by a geometric prior. This differs conceptually from CBAM or SGE: the mechanism does not merely infer saliency from features, but imposes a geometry-aware inductive bias inside the attention operator itself (Fan et al., 20 Apr 2026).

A fourth pattern is region-specific processing. S$^2$AM explicitly separates the spliced region from the background: $$y = M\times[L(G_{fg}(x)) + G_{mix}(x)] + (1-M)\times G_{bg}(x).$$ Here, the spliced region is processed through a learnable transformation block and dedicated gates, while the background is passed through a separate gate. The use of a Gaussian-smoothed mask indicates that the spatial prior is not merely binary segmentation, but a softened boundary prior meant to reduce seam artifacts (Cun et al., 2019).

A fifth pattern is cross-domain spatial refinement. TexLiverNet’s SFPM-f obtains coarse high-frequency information from skip-connected image features, applies FFT, enhances the frequency representation with a learnable filter, and reconstructs a refined feature: $$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$0

$$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$1

FSTA-SNN follows an analogous logic in spiking networks, but with DCT-based full-band spatial analysis rather than FFT-based decoder refinement (Jiang et al., 2024, Yu et al., 2024).

3. Architectural realizations across model families

In CNN backbones, spatial enhancement is often implemented as a plug-in residual or attention block. CBAM is explicitly described as lightweight and general, and can be integrated into any CNN architectures seamlessly with negligible overheads (Woo et al., 2018). SimAM in an enhanced ResNet50 for dental radiographs is inserted after the second ResNet block and augments the residual formulation as

$$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$2

with the stated aim of capturing spatial dependencies and enhancing significant features under low-contrast conditions (Rezaie et al., 2024). STEAM generalizes the channel-plus-spatial recipe by modeling both branches as graph attention problems: CIA builds a channel graph, while SIA constructs a spatial graph after Output Guided Pooling, which compresses an intermediate feature map into a fixed-size spatial map (Sabharwal et al., 2024).

Encoder-decoder systems frequently place the spatial module at low-level or skip-connected stages. S$$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$3AM is inserted in the coarser low-level features of a U-Net in two variants, S$$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$4ASC and S$$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$5AD, because harmonization is treated as primarily a low-level appearance correction problem (Cun et al., 2019). TexLiverNet places SFPM-s in a parallel structure to deeper layers and SFPM-f in the decoder/skip-connection stage; the first enriches deep spatial detail with standard and dilated convolutions, while the second uses adaptive frequency enhancement to clarify liver tumor boundaries (Jiang et al., 2024). In hyperspectral reconstruction, SEnet is not a single attention block but a reconstruction framework whose spatial enhancement path uses a grayscale measurement, an edge module, a multi-scale module, a deep-shallow fusion module, and an attention module to fuse spatial priors back into the hyperspectral estimate (Di et al., 29 Apr 2025).

Transformer and state-space architectures use more diverse spatial mechanisms. EVT’s EuSA sits after Conditional Positional Encoding and before the FFN: $$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$6 Its spatially-independent 1D grouping is used because spatial information is already supplied by the Euclidean decay matrix (Fan et al., 20 Apr 2026). MorpMamba’s “Spatial-Spectral Token Enhancement module” uses the center region of the HSI patch to gate both spatial and spectral tokens before multi-head self-attention and the state space model (Ahmad et al., 2024). STSMamba’s SDSpaM selects only the most informative spatial tokens, reorders them into a sparse deformable sequence, processes them with Mamba, and scatters the result back into the original spatial tensor as a residual update (Dewis et al., 29 Jul 2025). Remote sensing change detection pushes this logic further by scanning not only the $$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$7 plane but also $$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$8 and $$\mathbf{F'}=\mathbf{M_c}(\mathbf{F}) \otimes \mathbf{F}, \qquad \mathbf{F''}=\mathbf{M_s}(\mathbf{F'}) \otimes \mathbf{F'}.$$9 perspectives through 3D-SSM (Huang et al., 24 Jun 2025).

Spatial enhancement also appears in non-vision modalities. In target speech extraction, the spatial enhanced module is a multi-head cross-attention block inside neural beamforming, where spatial features derived from microphone geometry and target direction act as Query and covariance-matrix embeddings act as Key and Value (Guo et al., 2023). In optical mode decomposition, metasurface-enhanced spatial mode decomposition is not a neural block at all, but a Fourier-optics implementation that reduces cross-coupling through much finer spatial structuring than a spatial light modulator (Jones et al., 2021). These cases broaden the term beyond image attention and show that “spatial enhancement” can denote improved exploitation of geometry, location, or spatially structured measurements in a much wider systems sense.

4. Mathematical mechanisms

Despite their variety, most spatial enhanced modules can be grouped by the mathematical object they manipulate.

The first class manipulates a spatial mask. CBAM computes a single-channel attention map from pooled descriptors and applies element-wise multiplication (Woo et al., 2018). CSA-FE in remote sensing super-resolution uses a closely related spatial attention equation,

$7\times 7$0

so that important regions are highlighted before transformer processing (Sultan et al., 2024). STEAM’s SIA instead defines a graph $7\times 7$1 over a pooled $7\times 7$2 map, applies multi-head scaled dot-product graph attention, upsamples the resulting scores, and then modulates the channel-refined feature map (Sabharwal et al., 2024).

The second class manipulates global-local similarity. SGE measures each local descriptor against a group-level global descriptor and normalizes the response over spatial positions (Li et al., 2019). This is not the same as non-local attention: it compares each position to one semantic prototype rather than to every other position. A similar center-conditioned logic appears in MorpMamba, where the center token $7\times 7$3 produces gates for spatial and spectral tokens: $7\times 7$4

$7\times 7$5

The conditioning variable differs, but the principle remains local modulation guided by a compact global or center descriptor (Ahmad et al., 2024).

The third class manipulates explicit geometry-dependent weights. EVT attenuates token interactions with Euclidean distance decay (Fan et al., 20 Apr 2026). In beamforming, angle feature and inter-channel phase difference encode the geometry of the microphone array and target direction: $7\times 7$6 and these spatial cues guide cross-attention in the beamforming network (Guo et al., 2023). Geometry-aware modulation therefore need not be image-grid-based; it can arise from array geometry or token coordinates.

The fourth class manipulates frequency-domain spatial structure. FSTA-SNN computes

$7\times 7$7

interpreting global average pooling as only the $7\times 7$8 frequency component and using a fuller DCT spectrum to generate spatial weights (Yu et al., 2024). TexLiverNet uses FFT on coarse high-frequency features to separate useful high-frequency information from noise in decoder refinement (Jiang et al., 2024). In 3D-SSM change detection, the decoder’s MBFEM combines an FFT branch, a convolution branch, and a 3D-SSM branch so that boundary, local, and long-range spatial cues are represented jointly (Huang et al., 24 Jun 2025).

The fifth class manipulates sampling trajectories or token sequences. SG-DSCNet replaces fixed 3D convolution with Dynamic Snake Convolution, where kernel points are displaced by cumulative learned offsets: $7\times 7$9 so that the kernel follows object geometry rather than remaining on a rigid grid (Li et al., 6 Apr 2025). STSMamba accomplishes a related objective at the sequence level, not through deformed convolution but through sparse deformable token sequencing based on spatial attention to the center pixel (Dewis et al., 29 Jul 2025).

5. Integration strategies and reported effects

The empirical literature shows that spatial enhancement is rarely treated as an isolated replacement for the entire backbone. It is more often inserted at points where the base network is assumed to be losing locality, precision, or geometry. In CBAM, the module is general and end-to-end trainable with negligible overheads (Woo et al., 2018). In SGE, the component is inserted into CNN backbones with almost no extra parameters and calculations (Li et al., 2019). STEAM makes an even stronger efficiency claim: it is a constant-parameter module independent of the backbone, with a full ResNet-50 integration adding only 320 parameters and 3.57e-3 GFLOPs when $$\mathbf{M_s}(\mathbf{F})=\sigma\left(f^{7\times 7}\left([AvgPool(\mathbf{F}); MaxPool(\mathbf{F})]\right)\right).$$0 (Sabharwal et al., 2024).

Several papers report that these modules can materially improve performance without large complexity increases. SGE achieves 1.2\% Top-1 improvement on ImageNet with ResNet50 and 1.0$$\mathbf{M_s}(\mathbf{F})=\sigma\left(f^{7\times 7}\left([AvgPool(\mathbf{F}); MaxPool(\mathbf{F})]\right)\right).$$12.0\% AP gain on COCO detectors (Li et al., 2019). CBAM’s best spatial configuration on ResNet50 + channel + spatial reaches Top-1 Error 22.66\% and Top-5 Error 6.31\%, improving over channel-only and SE variants (Woo et al., 2018). EVT reports that on EVT-T, Euclidean decay gives 83.0 compared with 82.7 for Manhattan decay and 82.3 for no decay matrix, while the roadmap ablation attributes +0.4 top-1 and +1.2 mIoU to replacing Manhattan with Euclidean decay and an additional +0.6 top-1 and +1.9 mIoU to replacing decomposition with grouped EuSA (Fan et al., 20 Apr 2026). SCSC reports that FaceResNet with SCSC can improve 2.7\% with 68\% fewer FLOPs and 79\% fewer parameters, that Swin Transformer with SCSC can achieve better performance with 22\% fewer FLOPs, and that ResNet with SCSC can improve 5.3\% with similar complexity (Wang et al., 2023).

Task-specific modules report gains in more specialized metrics. TexLiverNet’s ablation on VTLiTS shows base+SFPM improving Dice from 80.39 to 81.36 and VOE from 28.53 to 25.92, while the joint base+TIA-Attention+SFPM reaches 82.74 Dice and 25.41 VOE (Jiang et al., 2024). FSTA reduces the overall spike firing rate by about 33.99\% while maintaining low energy consumption and small parameter overhead (Yu et al., 2024). In optical mode decomposition, the metasurface implementation measures a mode-weight fluctuation of $$\mathbf{M_s}(\mathbf{F})=\sigma\left(f^{7\times 7}\left([AvgPool(\mathbf{F}); MaxPool(\mathbf{F})]\right)\right).$$2 with 1 second of averaging at a Fourier frequency of 80 Hz, described as an improvement of more than three orders of magnitude over the state of the art of SLM-based decomposition (Jones et al., 2021). These results indicate that “spatial enhancement” is not confined to classification accuracy; it also targets power efficiency, firing sparsity, harmonization quality, spectral reconstruction, and measurement precision.

6. Conceptual boundaries, misconceptions, and open directions

A common misconception is to equate a spatial enhanced module with a generic spatial attention mask. The literature is broader. Some methods are indeed mask-based, such as CBAM or CSA-FE (Woo et al., 2018, Sultan et al., 2024). Others are explicitly geometry-biased, such as EVT’s Euclidean decay (Fan et al., 20 Apr 2026). Others remain spatial only in the sense of operating on regions or trajectories, as in S$$\mathbf{M_s}(\mathbf{F})=\sigma\left(f^{7\times 7}\left([AvgPool(\mathbf{F}); MaxPool(\mathbf{F})]\right)\right).$$3AM’s foreground-background separation or SG-DSCNet’s snake-like kernel deformation (Cun et al., 2019, Li et al., 6 Apr 2025). Still others are spatial-frequency modules, such as FSTA or SFPM-f, where spatial enhancement is mediated through DCT or FFT analysis rather than a direct spatial mask (Yu et al., 2024, Jiang et al., 2024).

Another misconception is that stronger spatial modeling necessarily requires heavy computation. Several papers argue the opposite. CBAM is described as lightweight, SGE uses almost no extra parameters and calculations, and STEAM is constant-parameter relative to the backbone (Woo et al., 2018, Li et al., 2019, Sabharwal et al., 2024). At the same time, efficiency claims are architecture-dependent. EVT reduces cost by replacing decomposed attention with 1D grouping because spatial information is already encoded in the decay matrix (Fan et al., 20 Apr 2026), whereas SE-RCASSI improves spatial quality partly by adding a second grayscale path in hardware, which is a different kind of system-level trade-off (Di et al., 29 Apr 2025).

The most plausible research direction is continued hybridization. Recent modules already combine spatial priors with token grouping, graph attention, frequency transforms, selective scan modeling, or external modalities such as text, grayscale images, and array geometry. This suggests that future “spatial enhanced modules” will likely be increasingly composite: neither purely convolutional nor purely attentional, neither purely spatial nor purely spectral, but mechanisms that allocate model capacity to location-sensitive structure while preserving efficiency, resolution adaptability, and compatibility with larger backbone families (Wang et al., 2023, Fan et al., 20 Apr 2026, Sabharwal et al., 2024, Huang et al., 24 Jun 2025).