Atrous Strip Convolution in DLO Segmentation
- Atrous strip convolution is a lightweight convolution design that combines dilation with strip operations to capture elongated, line-like features in image segmentation tasks.
- It expands the receptive field while reducing computational costs by employing parallel strip-shaped dilated branches and fusing decomposed results.
- Integrated within ASCSPP modules, it boosts segmentation accuracy for deformable linear objects (DLOs) while maintaining efficiency on edge computing platforms.
Atrous strip convolution is a lightweight hybrid convolutional design that combines strip convolution with atrous, or dilated, convolution in order to extract elongated, line-like features while retaining a large receptive field and low computational cost. In the current arXiv literature, the clearest explicit formulation appears in the ASC-SW framework for visual-assisted map navigation, where Atrous Strip Convolution (ASConv) is introduced inside the Atrous Strip Convolution Spatial Pyramid Pooling (ASCSPP) module of ASCnet to segment deformable linear objects (DLOs) such as wires and cables under edge-computing constraints (Liu et al., 17 Jul 2025). More loosely, the phrase can also refer to a family of related designs that combine dilation with directional, asymmetric, strip-like, or strip-aggregating operations, although several adjacent papers are careful not to claim a true strip-shaped atrous kernel as their primary contribution (Chee et al., 9 Feb 2026).
1. Emergence and problem setting
Atrous strip convolution emerged from a specific dense-prediction problem: segmentation of thin, long, deformable linear objects in mobile-robot navigation scenes. In ASC-SW, the motivating application is visual assistance for map-based mobile robot navigation using a depth camera and a lightweight visual neural network, with the explicit goal of compensating for the inability of traditional LiDAR sensors to detect ground-level obstacles such as ground-level wires (Liu et al., 17 Jul 2025). The operator is therefore not presented as a generic replacement for all convolutions, but as a task-specific mechanism for DLO perception under strict deployment budgets.
This framing distinguishes atrous strip convolution from earlier atrous-convolution research that was primarily concerned with receptive-field enlargement or degridding. For example, smoothed dilated convolution addresses the gridding artifact of dilated convolutions by inserting lightweight smoothing steps before or after the dilated operation, rather than redesigning the kernel into a strip geometry (Wang et al., 2018). A related distinction also appears in later segmentation work: DAS-SK combines dual atrous separable branches with strip pooling, but explicitly does not introduce strip-shaped atrous kernels as its main atrous operator (Chee et al., 9 Feb 2026). This suggests that atrous strip convolution, in the strict sense, should be reserved for methods whose receptive geometry is itself strip-based and dilated, rather than for any method that merely combines atrous processing with some strip-oriented context module.
2. Operator design and underlying rationale
ASConv is defined operationally rather than through a single closed-form equation. The ASC-SW paper states that “to better extract features of DLOs, we employ strip convolutions, and to meet the lightweight model requirement, we combine atrous convolutions with strip convolutions.” It further specifies that ASCnet does not rely on ordinary square kernels alone, but instead applies strip convolutions of the same size in parallel, varies the dilation rates to extract features, and fuses the decomposed results so as to capture higher-level linear features (Liu et al., 17 Jul 2025).
The design combines three ingredients. Standard convolution uses square kernels and dense local sampling, which is generic but not specialized for elongated structures. Atrous convolution enlarges the receptive field by sparsely sampling input positions while keeping the number of parameters and computational cost unchanged. Strip convolution replaces square kernels with 2D strip-shaped kernels, and vertical and horizontal strip convolutions can be cascaded to replace traditional square kernels with fewer parameters and computations. ASConv combines the latter two ideas: it preserves the strip-shaped receptive geometry that is sensitive to cables and wires, while using dilation to broaden the context seen by each branch without resorting to larger dense kernels (Liu et al., 17 Jul 2025).
A common misconception is to treat atrous strip convolution as merely an unusual kernel shape. In the ASC-SW formulation, that interpretation is incomplete. ASConv is described as a parallelized feature-extraction strategy that decomposes traditional convolution into multiple strip-based dilated branches and fuses them. The essential mechanism is therefore branch-wise decomposition and aggregation, not only anisotropic kernel selection. The paper also notes that dilation is added to strip convolution to reduce computational complexity while enlarging the receptive field, making the module especially suited to targets that have strong orientation structure and narrow spatial support (Liu et al., 17 Jul 2025).
3. Placement within ASCnet and ASCSPP
In ASC-SW, atrous strip convolution is not deployed as an isolated layer but as the core primitive inside ASCSPP, which replaces the standard ASPP block of DeepLabv3+. ASCnet follows a Deeplabv3+-style encoder-decoder architecture with MobileNetV2 as the backbone. The backbone output is split into two paths: a low-level feature path goes directly to the decoder, while a high-level feature path passes through ASCSPP for multiscale feature extraction (Liu et al., 17 Jul 2025).
ASCSPP modifies the atrous convolutions in ASPP by introducing atrous strip convolutions and using a parallel connection scheme. The paper emphasizes that this arrangement enables better feature extraction without increasing computational cost. It also adds the original low-level features before the output in order to compensate for spatial details lost in high-level features and to enrich the final representation (Liu et al., 17 Jul 2025).
This placement clarifies the architectural role of atrous strip convolution. It is not the encoder backbone itself, nor merely a decoder refinement stage. Rather, it serves as the multiscale contextual operator in the bottleneck-like pyramid module, where receptive-field shaping and directional bias are most consequential. The contrast with serial ASPP variants is explicit in the paper: modules such as SPASPP and DAPP enrich features by increasing the number of convolution operations and applying them in series, whereas ASCSPP uses a parallel connection scheme to increase feature richness without the same growth in parameters or compute (Liu et al., 17 Jul 2025).
4. Empirical characteristics and deployment profile
The ASC-SW study reports both task-level and module-level evidence for the utility of atrous strip convolution in DLO segmentation. At the system level, the framework achieves a mean Intersection over Union score of 75.3% on a self-built dataset and reaches 9.3 FPS inference speed on the Jetson Orin Nano edge device. The paper states that the overall approach outperforms existing DLO detection models and was validated on a physical robotic platform (Liu et al., 17 Jul 2025).
At the module level, the ablation tied to ASCSPP isolates the effect of the atrous strip design within the ASCnet framework.
| Module | mIoU | FPS |
|---|---|---|
| ASPP | 0.4386 | 223 |
| DAPP | 0.4906 | 205.92 |
| SPASPP | 0.336 | 205.47 |
| ASCSPP | 0.5069 | 217 |
These results show that ASCSPP achieved the best segmentation performance among the tested pyramid modules, while its inference speed was only slightly below plain ASPP on the reported 3070-device comparison (Liu et al., 17 Jul 2025). In practical terms, the gain over ASPP is about 6.8 percentage points absolute in mIoU, indicating that the strip-and-dilation combination improves task alignment more than it harms efficiency.
The reported performance profile also constrains interpretation. ASC-SW is lightweight and edge-deployable, but the paper explicitly notes that ASCnet still does not reach 30 FPS on edge hardware, so it is not fully real-time in the strictest sense (Liu et al., 17 Jul 2025). The evidence therefore supports a balanced claim: atrous strip convolution improves the accuracy-efficiency tradeoff for DLO segmentation, rather than eliminating the latency constraints of embedded inference.
5. Relation to neighboring atrous and strip-based designs
Atrous strip convolution belongs to a broader design space of receptive-field engineering, but adjacent methods differ materially in operator semantics. DAS-SK, for example, is directly concerned with atrous-separable design and enhanced ASPP, and it introduces six DAS-SKConv branches with dilation rates together with a strip pooling branch. However, the paper is explicit that the strip-like component lies in strip pooling rather than in strip-shaped atrous kernels, so the connection to atrous strip convolution is strong but indirect (Chee et al., 9 Feb 2026).
A similar distinction applies to the agricultural DAS-Conv model. That work replaces standard ASPP atrous branches with a dual-path module combining a standard atrous convolution and a depthwise atrous plus pointwise path. It is clearly an atrous-convolution innovation, but it does not propose a true strip convolution or anisotropic kernel design (Ling et al., 27 Jun 2025). In other words, dual atrous separable convolution expands the space of efficient dilated operators without introducing the geometry bias toward elongated structures that is central to ASConv.
Other related operators are even further from the strict notion. Kernel-sharing parallel atrous convolution for defocus deblurring uses multiple parallel square atrous branches with shared weights and learned scale and shape attention, but the paper explicitly states that the design is not strip-like or anisotropic (Son et al., 2021). Smoothed dilated convolution is instead a degridding framework based on separable and shared operations that smooth the dilated lattice rather than imposing a strip geometry (Wang et al., 2018). In image forensics, DMAC uses atrous convolution to preserve resolution and ASPP for multiscale localization, yet its relevance to atrous strip convolution remains conceptual rather than architectural (Liu et al., 2018).
The term has also been used loosely outside visual segmentation. In AcrE for knowledge graph embedding, atrous convolution is combined with standard 2D convolution and residual learning, and the phrase “atrous strip convolution” is used in a broad sense to describe sparse, widened filtering over a reshaped embedding grid. Yet the same description notes that the paper does not separately define the phrase as a distinct formal operator (Ren et al., 2020). In time-series forecasting, the “global atrous sliding window” and the central-bidirectional atrous algorithm are paired with semi-asymmetric convolution; this combination is directional and atrous, but it is a reconstruction-and-sampling procedure rather than the vision-oriented strip-kernel design found in ASC-SW (He, 2023).
6. Scope, limitations, and technical significance
The principal advantage of atrous strip convolution, as documented in ASC-SW, is its explicit orientation bias toward long, thin structures together with receptive-field enlargement through dilation. This improves sensitivity to the linear structure of DLOs, helps preserve mask continuity for narrow objects, and does so inside a lightweight architecture suited to edge deployment (Liu et al., 17 Jul 2025). A plausible implication is that the operator is most effective when target morphology is strongly elongated and when dense square-kernel context aggregation would be wasteful relative to the task geometry.
Its limitations are equally specific. The ASC-SW paper does not present a fully explicit mathematical derivation or an explicit kernel-size and dilation formula for ASConv itself, even though it gives a clear operational description. The performance gains are also tied to the DLO domain; the paper does not show that atrous strip convolution is universally superior across segmentation tasks (Liu et al., 17 Jul 2025). This is consistent with the broader literature, where neighboring methods often prefer strip pooling, dual atrous separable paths, or smoothing modules depending on whether the main challenge is directional structure, multi-scale efficiency, or degridding (Chee et al., 9 Feb 2026).
Taken together, the available arXiv evidence supports a narrow but technically coherent definition. Atrous strip convolution is best understood as a dilated, strip-oriented, multi-branch convolutional strategy for efficient extraction of elongated structures, with ASC-SW providing the most explicit current instantiation. Closely related methods illuminate the surrounding design space, but they also clarify the boundaries of the term: strip pooling is not the same as strip-shaped atrous kernels, dual atrous separable convolution is not inherently strip-based, and smoothing of dilated lattices is not a geometric strip operator. Within those boundaries, atrous strip convolution occupies a distinct niche at the intersection of receptive-field expansion, anisotropic feature bias, and low-cost dense prediction (Liu et al., 17 Jul 2025).
