MKRC: Multi-Kernel Residual Convolution

• MKRC is a neural network architecture that combines multi-scale convolutions and residual pathways to enhance feature extraction and performance in low-level computer vision tasks.
• It employs varying kernel sizes and dilation rates to capture both fine and coarse image details while maintaining low computational overhead.
• Empirical results in MRI reconstruction, image denoising, and dynamic convolution networks demonstrate MKRC's superior accuracy and parameter efficiency.

Multi-Kernel Residual Convolution (MKRC) is a family of neural network modules and architectures that leverage multiple effective convolutional kernel sizes in combination with residual pathways to enhance capacity, stability, and efficiency in low-level computer vision tasks. MKRC designs have been instantiated in diverse architectures for compressed sensing MRI reconstruction, image denoising, and dynamic convolution networks, unifying principles from multi-scale context aggregation, dilated convolutions, and residual learning (Dai et al., 2019, Yun et al., 2023, Liu et al., 2018).

1. Key Principles of MKRC

MKRC is defined by the integration of multiple receptive field sizes (achieved by varying kernel sizes or dilation rates) within a residual learning framework. It exploits the following core ideas:

• Multi-scale feature extraction: By employing convolutions at different scales—either through physically different kernel sizes (e.g., 3×3, 5×5, 7×7) or by using dilated convolutions with varying dilation rates (e.g., $d$ ∈ {1, 2, 3, 5})—MKRC modules capture fine-to-coarse context without incurring prohibitive parameter costs.
• Residual pathways: Both local and global residual skip connections are deployed. Local residuals preserve mid-level features and facilitate gradient flow within blocks, while global skips ensure that networks only need to model the correction to a base prediction.
• Efficiency: By combining multi-scale representation and residual connections, MKRC achieves high expressivity with controlled parameter and computational budgets. Dilated convolutions expand effective receptive fields using compact kernels, and kernel decomposition (static plus dynamic) enables adaptive processing with minimal overhead (Yun et al., 2023).

2. MKRC Variants and Architectural Realization

(a) Multi-Scale Dilated Residual Convolution Network for MRI

The architecture integrates multi-dilated 3×3 convolutions within blocks, alternating dilation rates between $d=3$ and $d=2$, with a final $d=1$ layer in each module. Each block employs a sequence of seven convolutions with alternately stacked channels (32 and 64), ReLU activations, five distinct local residual sums, and a block-level global residual skip. Two such blocks are chained and their outputs concatenated, followed by a fusion convolution to mix multi-scale features, and a final global residual from the input (Dai et al., 2019).

(b) Hybrid Dilated Convolution for Denoising

In the multiscale image denoising algorithm, the first layer performs parallel 3×3, 5×5, and 7×7 convolutions (concatenated), followed by groups of 3×3 dilated convolutions organized in a "Hybrid Dilated Convolution" (HDC) pattern, typically with dilation rates [1, 2, 5] to prevent gridding artifacts. Global and block-level residuals—with batch normalization and PReLU activation—are employed throughout (Liu et al., 2018).

(c) Dynamic Kernel-Space Residual Convolution

In the "Dynamic Mobile-Former" context, MKRC refers to the decomposition of a convolutional kernel $K$ as $$K(x) = K_{\text{static}} + \mathrm{diag}(\alpha(x)) K_{\text{dynamic}}$$, where $K_\text{static}$ is fixed across inputs and $K_\text{dynamic}$ is modulated per-sample using coefficients $\alpha(x)$ generated via a lightweight token-based transformer. This formulation anchors optimization with a static backbone, with dynamic refinements realized residually in kernel space (Yun et al., 2023).

3. Mathematical Formulation

(a) Dilated Convolution

Standard 2D convolution:

$$(F * x)(i) = \sum_{k \in \Omega} F(k) \cdot x(i - k)$$

Dilated ($d=3$0-dilated) convolution: $d=3$1 For $d=3$2 kernel and $d=3$3, receptive field equates to $d=3$4, but only retains nine parameters.

(b) MKRC Kernel-Space Residual

$d=3$5

with $d=3$6 predicted from a global feature summary via a transformer and small MLP branch (Yun et al., 2023).

4. Empirical Results and Ablation Findings

Quantitative evaluations demonstrate that MKRC-enabled architectures achieve strong performance with parameter efficiency and robustness:

Task	Baseline	MKRC Variant	PSNR (dB)	SSIM	Speed/Params
CSMRI (random 20%)	2-block non-dilated	2-block MKRC	34.95	0.944	-30% params
Denoising (BSD68, σ=25)	DnCNN	MKRC (HDC)	29.26	-	~0.5x params
ImageNet (B1)	Mobile-Former	Mobile-Former+MKRC	74.2% (Top-1)	-	+1.7% FLOPs

Key ablations (Dai et al., 2019, Yun et al., 2023, Liu et al., 2018):

• Multi-dilated or multi-kernel designs outperform single-scale or non-dilated variants, especially when parameter-matched.
• Local and global residuals must both be present for maximal performance; removing either component degrades PSNR/SSIM by up to 1 dB.
• Concatenating multi-block (multi-scale) features, rather than summing or omitting, significantly improves metrics.

5. Applications and Practical Considerations

Compressed Sensing MRI

The MDN architecture employing MKRC yields superior reconstruction accuracy and image edge detail versus classical compressed sensing and other deep models. Robustness to k-space noise is empirically verified, with MKRC outperforming comparable architectures by 0.5–0.7 dB (Dai et al., 2019). Extension to MRI super-resolution tasks further demonstrates adaptability.

Image Denoising

MKRC achieves PSNR parity with or improvement over DnCNN and BM3D, while requiring approximately half the parameters and reduced inference time. The design is suitable for both grayscale and color images, and addresses gridding artifacts via hybrid dilation scheduling (Liu et al., 2018).

Dynamic Convolutional Networks

In dynamic convolution settings, the kernel-space residual formulation allows adaptive filtering per input with a marginal computational overhead (<5% of baseline), delivering 0.5–0.8% absolute accuracy gain on ImageNet for mid-sized models (Yun et al., 2023).

6. Design Rationale and Theoretical Implications

MKRC frameworks are constructed to maximize representational power under resource constraints:

• Dilated or multi-kernel convolutions provide large receptive fields with minimal parameter increase.
• Residual connections stabilize training, facilitate deeper architectures, and mitigate over-smoothing of high-frequency content.
• Multi-path concatenation aggregates features across scales, enhancing both low- and high-level processing within a lightweight footprint.
• Kernel-space residualization in dynamic models decouples static representational strength from dynamic adaptation, enabling stable optimization.

A plausible implication is that such architectures will generalize well to inductive bias-critical domains (e.g., medical imaging, low-level vision) and can be extended to dynamic, input-conditioned adaptation in more general deep nets.

7. Limitations and Considerations

While MKRC approaches yield state-of-the-art performance with efficiency, several constraints are noted:

• Manual tuning of dilation rates, residual structures, and kernel sharing are required for optimal results and may not transfer universally.
• The effectiveness of channel-wise dynamic modulation may be limited in highly heterogeneous input domains; more granular (spatial or sub-channel) adaptations incur higher computation.
• Performance gain saturates beyond a certain number of dilated/multi-kernel stages; excess multi-scale depth can lead to diminishing returns or redundant feature extraction.

Experimental evidence across diverse MKRC instantiations confirms the general effectiveness and versatility of the multi-kernel residual paradigm (Dai et al., 2019, Yun et al., 2023, Liu et al., 2018).