Large Strip Convolutions in Neural Networks

• Large Strip Convolutions are a technique that decomposes square kernels into sequential horizontal and vertical filters, enabling efficient anisotropic representation.
• They reduce parameter and computational costs significantly while maintaining an effective large receptive field for detecting elongated structures.
• Empirical evaluations on models like StripRFNet and Strip R-CNN demonstrate improved detection metrics in applications such as remote sensing and crack detection.

Large strip convolutions comprise a class of convolutional operators in neural networks and harmonic analysis designed for efficient long-range feature extraction along particular axes. In modern deep learning, these operators are typically implemented as the sequential application of horizontal and vertical one-dimensional (1×K and K×1) convolutional filters, providing an anisotropic expansion of the receptive field at significantly lower computational and parameter cost than conventional large square kernels. In harmonic mapping theory, “strip mappings” have a distinct analytic definition, with convolution (Hadamard product) operations producing convexity in specific directions. The concept is central in state-of-the-art models for object detection and semantic segmentation tasks where elongated patterns (e.g., cracks, vessels, text, high-aspect-ratio structures) dominate.

1. Mathematical Formulation and Definition

A large strip convolution is formally realized by decomposing a square convolutional kernel of size $k \times k$ into two successive one-dimensional kernels: a horizontal $1 \times K$ convolution followed by a vertical $K \times 1$ convolution, typically with kernel size $K \gg 1$. For an input tensor $X \in \mathbb{R}^{C \times H \times W}$ (channels $C$, height $H$, width $W$), the transformation comprises:

• Initial local context extraction via depthwise $k \times k$ convolution: $F_{sq} = \mathrm{DWConv}_{k\times k}(X)$
• Horizontal strip convolution: $F_h = \mathrm{Conv}_{1 \times K}(F_{sq})$
• Vertical strip convolution: $F_v = \mathrm{Conv}_{K \times 1}(F_h)$

These are followed by pointwise convolutions for channel mixing and gating. Optionally, dilation can be introduced to further expand the receptive field, as in $\mathrm{Conv}_{1 \times (K + 2(d-1)),d}$.

Parameter and computational cost analysis reveals significant savings: a $k \times k$ kernel involves $O(k^2)$ parameters per channel, while two strip convolutions require only $O(2k)$ per channel, yielding over an order of magnitude reduction for large $k$. The nominal receptive field is retained as $K \times K$, but the effective coverage is “cross-shaped,” focusing model capacity on elongated structures (Lin et al., 17 Oct 2025, Wang et al., 7 Sep 2025, Yuan et al., 7 Jan 2025).

2. Architectural Integration and Variants

Large strip convolutions are frequently embedded in specialized modules. Examples include:

• Strip Receptive Field Module (SRFM) in StripRFNet (Lin et al., 17 Oct 2025): integrates three sequential depthwise sub-blocks, replacing standard bottlenecks with cascaded strip convolutions and global strip pooling.
• Strip Attention Block (SAB) in StripDet (Wang et al., 7 Sep 2025): combines strip convolutions with spatial attention maps, enabling long-range context aggregation in lightweight 3D object detectors.
• StripNet backbone in Strip R-CNN (Yuan et al., 7 Jan 2025): each block features a local $5 \times 5$ depthwise convolution, followed by strip convolutions, pointwise channel mixing, and residual connections.

Strip-based heads for detection tasks leverage strip convolutions for improved localization of elongated objects. In heads, as in Strip R-CNN, localization branches incorporate strip modules that consist of standard and strip convolutions to regress bounding box coordinates efficiently. Typical kernel sizes (e.g., $K = 19$) are empirically validated for remote sensing patches of size $1024\times 1024$ (Yuan et al., 7 Jan 2025).

3. Empirical Evaluation and Efficiency

Ablation studies consistently demonstrate the efficacy of large strip convolutions for high-aspect-ratio targets:

• On RDD2022, StripRFNet with SRFM achieves F1 of 81.6% (+4.4 percentage points over YOLO11s baseline), mAP$50$ of 84.4% (+2.9 pp), and mAP$50:95$ of 52.0% (+3.4 pp) (Lin et al., 17 Oct 2025).
• StripDet yields car-3D moderate mAP of 78.1% on KITTI with only 0.65M parameters versus 75.9% for PointPillars (4.83M params), cutting parameters by $\sim 7\times$ (Wang et al., 7 Sep 2025).
• Strip R-CNN achieves 82.75% mAP on DOTA-v1.0 (multi-scale), outperforming LSKNet-S and PKINet-S (Yuan et al., 7 Jan 2025).

Computational metrics further underline their efficiency:

• SRFM reduces parameter count by 5.7% and FLOPs by 5.8% relative to P2-only StripRFNet (Lin et al., 17 Oct 2025).
• StripDet’s backbone achieves linear scaling in $k$ ($O(HWk)$) versus quadratic for square kernels ($O(HWk^2)$), rendering large receptive field extraction feasible in mobile and edge scenarios (Wang et al., 7 Sep 2025).
• Indirect convolution implementation of 1×N or N×1 kernels, as described in (Dukhan, 2019), provides 3–5× throughput boost over direct convolution with up to 90% reduction in temporary buffer size for large N.

Model/Module	Kernel Size	Params (M)	FLOPs (G)	mAP (%)	Throughput	Latency/Speed
StripRFNet (SRFM)	1×K, K×1	9.95	26.0	81.6	80 FPS	12.0 ms/img
StripDet (SAB)	1×K, K×1	0.65	9.5	79.97	N/A	N/A
Strip R-CNN (K=19)	1×K, K×1	30.5	159	82.75	17.7 FPS	N/A

Kernel size $K=19$ typically yields the best results for elongated object detection in aerial imagery (Yuan et al., 7 Jan 2025).

4. Implementation Considerations and Optimization

Efficient realization of large strip convolutions is critically dependent on hardware-aware strategies:

• Indirect Convolution Algorithm (Dukhan, 2019) replaces the traditional im2col approach, using pointer indirection buffers to minimize memory movement for large strip kernels. The indirection buffer (size $M \times R'$ pointers) greatly reduces overhead relative to im2col patch matrices ($O(CMHW)$ floats).
• Arithmetic intensity is doubled versus im2col for large input channels, improving cache utilization.
• On CPUs, cache blocking and pointer alignment optimize access; vectorization (AVX2/AVX-512, NEON) is imperative for maximizing throughput.
• On GPUs, shared-memory tiling and warp-cooperative reads accelerate inner loops, with the indirection buffer accessed from global memory and filters cached read-only.

Benchmarks for strip kernel lengths $N = 32, 128, 512$ demonstrate parity or respective gains over GEMM+im2col with a fraction of the memory usage, and accuracy remains within $<1$ ULP for all methods (Dukhan, 2019).

Strip Length N	Method	GFLOPS	im2col Buf (MB)	indir Buf (MB)
32	Direct conv	12.3	0.00	0.00
	GEMM+im2col	48.5	0.11	0.00
	Indirect GEMM	51.0	0.00	0.003
128	Direct conv	3.4	0.00	0.00
	GEMM+im2col	14.8	0.44	0.00
	Indirect GEMM	15.6	0.00	0.012

5. Application Domains and Design Principles

Large strip convolutions are optimally suited to problems characterized by spatial anisotropy:

• Crack detection: Road cracks with high aspect ratios benefit from receptive fields tuned to the axis of elongation (Lin et al., 17 Oct 2025).
• Remote-sensing object detection: Objects such as ships, roads, runways, and text lines display natural orientation, making strip convolution-based backbones preferable (Yuan et al., 7 Jan 2025).
• 3D point cloud detection: Directional dependencies in bird’s-eye-view features are efficiently modeled via strip attention (Wang et al., 7 Sep 2025).

Guidelines for deployment include:

• Use sequential $1 \times K$ and $K \times 1$ depthwise convolutions for parameter efficiency.
• Employ strip pooling or attention mechanisms for global context fusion along major axes.
• Always combine strip convolutions with pointwise layers and residual connections for stable training.
• Maintain kernel size across backbone stages or adjust based on input resolution to ensure consistent receptive field growth.

Limitations are context-dependent:

• Purely horizontal/vertical anisotropy is addressed; diagonal or fully isotropic patterns may require additional kernels.
• For small $K$, efficiency gains are minimal; for very large $K$, hardware bandwidth and latency must be considered (Wang et al., 7 Sep 2025).

6. Strip Mappings and Harmonic Convolutions

In the harmonic analysis context, a strip mapping is an analytic or harmonic map from the unit disk $D$ to a horizontal or slanted strip in $\mathbb{C}$, defined as $f=h+\overline{g}$, where $h(z)$, $g(z)$ are analytic. The convolution (Hadamard product) of two such mappings, $f_1★f_2$, defined by coefficientwise multiplication, preserves or produces convexity in a specified direction.

Key results include:

• Convolution of a half-plane mapping $f_1 \in S(H_{\theta,a})$ with a strip mapping $f_2 \in S(\Pi_{\varphi,b})$ yields a mapping convex in direction $-(\theta+\arg(1+a)+\arg(1+b))$, provided local univalence (Li et al., 2019).
• Convolution structure carries over under rotation; thus strip-to-strip convolutions inherit convexity properties from half-plane convolutions.
• Examples show analytic calculation of ranges and convexity directions for composed mappings.

7. Practical Themes and Future Directions

Large strip convolutional designs increasingly permeate architectures focused on edge deployment, anisotropic object detection, and domains with structured spatial patterns. Empirical successes and rigorous theoretical formulation demonstrate their utility across modalities, provided kernels are sized and implemented subject to hardware and data constraints.

A plausible implication is the emergence of architectures combining multiple directional convolutions (potentially diagonal or rotated strips) or combining strip-based operators with standard square, circular, or dilated kernels to further enhance spatial representation power. This suggests ongoing research will likely investigate data-driven kernel composition strategies adapting local receptive field shapes to target distributions.

References:

StripRFNet (Lin et al., 17 Oct 2025), StripDet (Wang et al., 7 Sep 2025), Strip R-CNN (Yuan et al., 7 Jan 2025), Indirect Convolution Algorithm (Dukhan, 2019), Rotations and convolutions of harmonic convex mappings (Li et al., 2019).