Enhanced Multi-Scale Dilated Convolution (E-MSDC)

• E-MSDC is a multi-scale convolution framework that combines varied dilation rates, stacking, and parallel branches to expand receptive fields while preserving spatial resolution.
• It integrates fusion strategies such as concatenation, residual addition, and adaptive weighting to effectively aggregate features across different scales.
• Empirical studies demonstrate that E-MSDC enhances performance in tasks like MRI reconstruction, human pose estimation, and time series forecasting by reducing errors and computational costs.

Enhanced Multi-Scale Dilated Convolution (E-MSDC) denotes, in the literature summarized here, a family of convolutional designs that expands receptive fields by combining multiple dilation rates and then fuses the resulting features through concatenation, residual addition, weighted aggregation, or attention-like recalibration. One comparative account refers to E-MSDC as “Explicit Multi-Scale Dilated Conv”, characterized by manual or fixed mixes of dilation rates per convolution; later work extends the same design space with adaptive, searched, deformable, dense, residual, and equilibrium-based variants (Liu et al., 2020, Luo et al., 2021, Yu et al., 2024, Li et al., 9 Apr 2026). This suggests that E-MSDC is best treated not as a single canonical layer, but as a recurrent architectural principle for multi-scale context aggregation under parameter and resolution constraints.

1. Foundational mechanism and formal structure

The modern lineage of E-MSDC begins with the observation that dilated convolutions can enlarge receptive fields without loss of resolution or coverage. In the standard formulation, a dilated convolution with dilation factor $l$ is written as

$$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$

which reduces to ordinary convolution when $l=1$. A stacked schedule of increasing dilations,

$F_{i+1} = F_i *_{2^i} k_i,$

produces exponential receptive-field growth while preserving spatial resolution. A representative context module uses 7 layers of $3\times3$ convolutions with dilation factors 1, 1, 2, 4, 8, 16, 1, followed by a $1\times1$ convolution, and maintains identical input and output shape so that it can be inserted into dense prediction systems without pooling or subsampling (Yu et al., 2015).

A second foundational pattern replaces large-kernel branching with multiple branches that share kernel size but differ in dilation. In single-image super-resolution, a dilated-convolution-based inception module applies three parallel $3\times3$ convolutions with dilation factors 1, 2, and 3, then concatenates the outputs:

$$y = \mathrm{Concat}\left(f_{l=1}(x),\ f_{l=2}(x),\ f_{l=3}(x)\right).$$

Cascading such modules yields a deeper network whose effective receptive field grows while parameter growth remains controlled (Shi et al., 2017).

These two patterns—stacked dilation schedules and parallel multi-dilation branches—form the core algebra of E-MSDC. The first emphasizes progressive context expansion; the second emphasizes simultaneous multi-scale feature capture. Much of the later literature consists of different ways of fusing, regularizing, or adapting these two constructions.

2. Recurring architectural motifs

E-MSDC systems recur around a small set of architectural motifs. One of the clearest is the combination of multi-scale dilation with residual and concatenation pathways. In compressed sensing MRI, a multi-scale dilated residual convolution network uses an MDN block with seven convolutional layers: 1 standard convolution layer, 3 layers with dilation rate 2, and 3 layers with dilation rate 3. The design further adds global residual learning, local residual learnings, and concatenation layers to fuse multi-scale features and preserve image edges and details (Dai et al., 2019).

A second motif is parallel long-range and short-range decomposition. In long-term time series forecasting, MSDCN employs two parallel convolutional modules: a long-term module with a larger kernel size and a short-term module with a smaller kernel size. Within each module, several parallel 1D dilated convolution blocks use exponentially growing dilation factors and are fused with learnable channel-wise weights:

$M = \sum_{i=1}^{n+m} H_i \odot W'_i.$

This fused nonlinear component is then combined with a linear autoregressive module, producing a hybrid model that explicitly joins nonlinear multi-scale extraction with linear temporal dependency modeling (Li et al., 2024).

A third motif is artifact-aware dilation scheduling. In image denoising, a lightweight residual model combines a multi-scale convolution group with hybrid dilated convolution (HDC). The HDC portion uses dilation rates 1, 2, 5 to enlarge the receptive field while reducing gridding artifacts; residual connections and batch normalization stabilize optimization, and the network predicts the noise residual rather than the clean image directly (Liu et al., 2018).

A fourth motif is parallel receptive-field ensembles with dense aggregation. In MRI reconstruction, an ensemble denoiser block processes branches with effective receptive fields of $3\times3$, $$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$0, $$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$1, and $$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$2, while dense residual connections operate both within branches and across the cascade. The resulting architecture preserves resolution, aggregates large structural artefacts and fine local detail simultaneously, and uses a deep cascade global architecture to reduce overfitting (Ma et al., 2023).

Taken together, these designs suggest that the “enhanced” aspect of E-MSDC usually lies less in dilation alone than in the surrounding fusion policy, residual topology, and artifact-control strategy.

3. Adaptive, searched, and deformable variants

A major limitation of early E-MSDC designs is that dilation rates are typically fixed. Several later systems explicitly challenge that assumption.

Adaptive Dilated Convolution (ADC) for human pose estimation assigns a distinct, learned dilation rate to each channel or channel group, with rates produced by a Dilation-Rates Regression Module (DRM) composed of global average pooling and two fully connected layers. ADC also supports fractional dilation rates, with bilinear interpolation used at sampling points. Because all multi-scale features are generated from the same-size feature map, spatial alignment is preserved while scale adaptivity is introduced (Luo et al., 2021).

ASCNet pushes this further by learning a per-pixel dilation rate field. Its rate-prediction subnetwork is a compact 3-layer convolutional module with channel numbers 8, 4, and 1, and the adaptive-scale convolution is defined as

$$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$3

where $$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$4 is a learned, floating-point, spatially varying dilation rate. Bilinear interpolation again handles non-integer sampling locations (Zhang et al., 2019).

A different branch of work treats dilation as a search variable. Inception convolution allows independent dilation patterns across spatial axes, output channels, and layers, so each output channel may carry its own pair $$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$5. Efficient Dilation Optimization (EDO) then selects patterns by minimizing the $$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$6 difference between full supernet weights and dilation-masked weights. In its own comparison, this work characterizes E-MSDC as using manual or fixed mixes of dilation rates per convolution, contrasting that with a channel-wise and axis-wise learned alternative (Liu et al., 2020).

Deformable multi-scale variants add geometric adaptivity. In traffic object detection, the MSDDC block first predicts offsets,

$$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$7

then applies three parallel deformable convolutions with dilation rates $$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$8,

$$(F *_{l} k)(\mathbf{p}) = \sum_{\mathbf{s} + l\mathbf{t} = \mathbf{p}} F(\mathbf{s})\, k(\mathbf{t}),$$9

and fuses them through concatenation and a $l=1$0 convolution:

$l=1$1

This design combines multi-scale dilation and learned sampling geometry, rather than treating them as separate mechanisms (Li et al., 9 Apr 2026).

These extensions collectively indicate a broad shift from static multi-scale context aggregation toward input-conditioned receptive-field control.

4. Empirical record across application domains

The empirical literature on E-MSDC spans dense prediction, restoration, inverse problems, sequence modeling, and audio. The diversity of tasks is notable because the same underlying design principle—multiple receptive fields fused at constant or near-constant computational budget—appears repeatedly in otherwise unrelated settings.

Domain	Reported outcome	Paper
Long-term time series forecasting	2.4% reduction in MSE and 2.2% reduction in MAE vs. TimesNet; 9.4% reduction in MSE and 11.0% reduction in MAE vs. ETSformer	(Li et al., 2024)
Human pose estimation	AP improves from 70.4 to 71.8 for SimpleBaseline-Res50, and from 75.1 to 75.5 for HRNet-W48	(Luo et al., 2021)
Image compressed sensing	MsDCNN-3: 26.43 dB at MR=0.10; at MR=0.01, 20.22 dB vs. 11.31–17.54 dB for comparators	(Wang et al., 2022)
MRI reconstruction	The real-valued model outperformed common reconstruction architectures as well as a state-of-the-art multi-scale network whilst being three times more efficient	(Ma et al., 2023)
Speech separation	SI-SDRi = 20.3 dB, SDRi = 20.5 dB, 6.14M parameters, 1.81 s CPU inference for a 4-second sample	(Wang, 2023)
Traffic object detection	Hybrid backbone reaches 92.1 mAP@50 and 69.7 mAP@50-95 with 5.0M parameters and 11.6 FLOPs	(Li et al., 9 Apr 2026)

Outside the table, the same pattern appears in additional domains. In blur detection, MSDU-net uses a group of multi-scale feature extractors with dilated convolutions and a U-shape fusion path, outperforming other state-of-the-art blur detection methods on two publicly available benchmarks (Yang et al., 2020). In singer identification, waveform-domain MDCNN-SID uses multi-scale dilated convolutions and skip connections, reaching 0.854 song-level accuracy on Artist20 and 0.796 on Singer107 (Zhang et al., 2020). In domestic activity classification, a dilated depthwise separable network with multi-scale embedding attains 0.831 classification accuracy, outperforming lightweight baselines such as MobileNet-V1, MobileNet-V2, and ShuffleNet (Zeng et al., 2023).

This breadth suggests that E-MSDC is less a domain-specific trick than a broadly reusable inductive bias for settings where both local detail and nonlocal context matter.

5. Limitations, failure modes, and common misconceptions

A common misconception is that increasing dilation automatically improves multi-scale modeling. The literature does not support such a simple conclusion. In MsDC-DEQ-Net, branch 1 (df=1) is described as crucial, branch 2 (df=2) as also significant, and removing higher-dilation branches such as 3 and 6 has little effect; the best result is achieved when all branches are active, but the marginal utility of large dilation is clearly non-uniform (Yu et al., 2024).

A second misconception is that any multi-scale fusion preserves spatial consistency. Traditional human pose estimation pipelines often fuse feature maps at $l=1$2, $l=1$3, $l=1$4, and $l=1$5 resolution, but this strategy is described as suffering from spatial misalignment and lack of scale adaptivity. ADC addresses this by generating and fusing multi-scale features at the same spatial size, using different dilation rates for different channels (Luo et al., 2021).

A third misconception is that dilated convolution is free of sampling artefacts. Two independent lines of work explicitly identify the gridding effect as a failure mode. In image compressed sensing, stacking several dilated convolutions with the same dilation factor is said to cause gridding and loss of continuity, which motivates alternating dilated and standard convolutions within each channel (Wang et al., 2022). In image denoising, the hybrid schedule $l=1$6 is introduced specifically to decrease gridding artifacts relative to plain repeated dilation (Liu et al., 2018).

These results imply that E-MSDC performance depends critically on how scales are fused, whether alignment is preserved, and how dilation schedules are regularized. Dilation enlarges coverage, but it does not by itself solve scale selection, feature alignment, or sampling continuity.

6. Relation to adjacent paradigms and likely directions

The design logic of E-MSDC has increasingly migrated beyond conventional convolution. In visual recognition, DilateFormer introduces Multi-Scale Dilated Attention (MSDA) to model local and sparse patch interaction within sliding windows, then stacks MSDA blocks at low-level stages and global multi-head self-attention blocks at high-level stages. On ImageNet-1K classification, DilateFormer reports comparable performance with 70% fewer FLOPs than existing state-of-the-art models, indicating that the multi-scale dilated principle can be transferred from convolutional sampling to attention patterns (Jiao et al., 2023).

At the same time, E-MSDC modules are being embedded inside broader hybrid systems. In speech separation, ARFDCN couples dilated convolutions with recurrent fusion across stages and a channel attention module, using the recurrence

$l=1$7

In compressive sensing, MsDC-DEQ-Net inserts multi-scale dilated convolutions into an ISTA-mapped deep equilibrium block, together with ResNeXt and SE mechanisms, and uses a single iteration block that can be iteratively reused (Wang, 2023, Yu et al., 2024). In traffic detection, MDDCNet combines shallow-stage MSDDC blocks with deeper Mamba blocks, adds a Channel-Enhanced Feed-Forward Network (CE-FFN), and constructs a Mamba-based Attention-Aggregating Feature Pyramid Network (A^2FPN) for stronger cross-scale interaction (Li et al., 9 Apr 2026).

A plausible implication is that the future of E-MSDC lies in hybridization rather than isolation: same-resolution fusion to preserve alignment, adaptive or searched dilation to avoid static scale bias, deformable sampling for geometric variability, and integration with attention, state-space, or equilibrium formulations when global modeling is required. Under that reading, E-MSDC remains a central local-context operator, but no longer a self-sufficient architecture.