ASCnet: Atrous Strip Convolution for DLO Segmentation

• ASCnet is a lightweight segmentation network that uses atrous strip convolution to detect deformable linear objects in challenging robotic environments.
• It integrates a MobileNetV2 backbone with a DeepLabv3+-style encoder-decoder and replaces conventional ASPP with the efficient ASCSPP module.
• The network enhances segmentation stability through a sliding window post-processing module, significantly improving mIoU for low-profile obstacle detection.

Atrous Strip Convolution Network (ASCnet) is a lightweight deformable linear object (DLO) segmentation network introduced within the visual-assisted navigation framework "ASC-SW: Atrous strip convolution network with sliding windows for visual-assisted map navigation" (Liu et al., 17 Jul 2025). It is designed for map-based mobile robot navigation in settings where LiDAR often fails to detect low-profile, ground-level obstacles such as wires and cables. ASCnet uses MobileNetV2 as its backbone, adopts a DeepLabv3+-style encoder–decoder, and replaces conventional atrous spatial pyramid pooling with Atrous Strip Convolution Spatial Pyramid Pooling (ASCSPP) to capture thin, elongated, piecewise-linear structures at low computational cost. In the same framework, a Sliding Window (SW) post-processing module suppresses flickering false detections through temporal voting on tracked connected components, and the resulting mask is used to filter depth measurements and insert DLO point clouds into the navigation costmap (Liu et al., 17 Jul 2025).

1. Terminology, naming, and scope

The term ASCnet in (Liu et al., 17 Jul 2025) denotes Atrous Strip Convolution Network, a semantic segmentation model for DLO detection in RGB images acquired for mobile robot navigation. Its principal application context is the ASC-SW visual-assisted navigation pipeline, where RGB-D data from a depth camera are combined with lightweight segmentation and temporal post-processing to identify obstacles that are difficult for LiDAR to perceive, especially ground-level wires (Liu et al., 17 Jul 2025).

A recurrent source of ambiguity is that ASCNet has also been used to mean Asymmetric Sampling Correction Network in infrared image destriping research (Yuan et al., 2024). That model addresses column non-uniformity and stripe noise in infrared imagery and explicitly does not use atrous or dilated convolutions (Yuan et al., 2024). In contrast, the ASCnet of (Liu et al., 17 Jul 2025) is a DLO segmentation architecture centered on atrous strip convolution, ASCSPP, and deployment in visual-assisted robotic navigation. The shared acronym therefore refers to two distinct architectures in different subfields.

This distinction is methodologically important. The navigation-oriented ASCnet is a fully convolutional RGB segmentation network following the DeepLabv3+ paradigm, whereas the infrared destriping ASCNet is a U-shaped restoration model built around Residual Haar Discrete Wavelet Transform and Column Non-uniformity Correction Module components (Yuan et al., 2024). Any discussion of Atrous Strip Convolution Network in the context of DLO detection, ASCSPP, or ASC-SW properly refers to the 2025 navigation paper (Liu et al., 17 Jul 2025).

2. Problem setting and role in visual-assisted navigation

ASCnet was motivated by a specific sensing gap in mobile robotics: most systems rely on LiDAR for obstacle sensing, yet LiDAR often fails to detect low-profile, ground-level obstacles such as wires and cables. The paper frames these objects as deformable linear objects (DLOs) and identifies them as a source of safety and navigation failures in map-based mobile robot navigation (Liu et al., 17 Jul 2025).

The detection problem is difficult for several reasons. DLOs are thin, elongated, deformable, and may appear under diverse viewpoints. They often share colors with the background, including black, white, or yellow wires against similarly colored floors. The paper further notes that many DLO segmentation datasets are captured from a top-down viewpoint perpendicular to the plane, whereas a navigating robot typically observes the scene from an oblique viewpoint. This mismatch contributes to false positives and false negatives in real scenes (Liu et al., 17 Jul 2025).

Within the broader ASC-SW framework, ASCnet is the per-frame segmentation core. The pipeline is organized as follows. RGB-D data are obtained from a depth camera, specifically an Intel RealSense D455 in the reported tests. ASCnet segments DLOs from RGB images. The SW module then denoises the mask through temporal voting on tracked connected components to suppress flickering false detections in complex environments. The cleaned mask is used to filter the depth map and extract the DLO point cloud, voxel downsampling reduces computation, and the filtered point cloud is inserted into the navigation costmap as obstacles (Liu et al., 17 Jul 2025).

This arrangement indicates that ASCnet is not merely a standalone segmentation model but part of a perception-to-planning chain. A plausible implication is that its design priorities—low computational load, compatibility with edge devices, and robustness to oblique viewpoints—are shaped by downstream navigation constraints rather than by segmentation accuracy alone.

3. Network architecture

ASCnet is described as an encoder–decoder network based on DeepLabv3+ with a MobileNetV2 backbone and several targeted modifications for DLO segmentation (Liu et al., 17 Jul 2025). MobileNetV2 provides inverted residuals and depthwise separable convolution for efficient feature extraction. Its output is split into two branches: low-level features are routed directly to the decoder for spatial detail recovery, while high-level features are sent to ASCSPP for multi-scale context extraction (Liu et al., 17 Jul 2025).

The principal architectural additions are threefold. First, the encoder head replaces classic ASPP with Atrous Strip Convolution Spatial Pyramid Pooling (ASCSPP). Second, the model includes a lightweight channel attention (CA) module, described as a squeeze-and-excitation variant that uses a $1 \times 1$ convolution instead of FC+ReLU. Third, the decoder follows the DeepLabv3+ fusion strategy, combining high-level features from ASCSPP with low-level features from the backbone via a skip connection to recover spatial details (Liu et al., 17 Jul 2025).

The paper does not enumerate the exact upsampling layer types, strides, padding, or final classifier head details, but it explicitly characterizes the model as a fully convolutional DeepLabv3+-style encoder–decoder (Liu et al., 17 Jul 2025). The CA module operates by applying global pooling to produce channel descriptors, using a $1 \times 1$ convolution to generate channel weights, and multiplying those weights channel-wise with the feature map. The modification is reported to reduce compute relative to classic SE while improving accuracy (Liu et al., 17 Jul 2025).

The paper reports the following model sizes and compute footprints:

Model	Parameters	GFLOPs
ASCnet	6.9M	6.36
DeepLabV3Plus-MobileNetV2	6.8M	6.34
DeepLabV3Plus-ResNet101	59.9M	14.53
RT-DLO (640×360)	45.6M	50.18

These numbers place ASCnet very close to DeepLabV3Plus-MobileNetV2 in compute, while adding the DLO-specific ASCSPP and CA mechanisms (Liu et al., 17 Jul 2025). This suggests that the architectural modifications were intended to preserve the deployment profile of a lightweight segmentation backbone while improving linear-structure sensitivity.

4. Atrous Strip Convolution and ASCSPP

The core operator in ASCnet is Atrous Strip Convolution (ASConv). The paper begins from the standard 2D convolution

$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$

and the dilated convolution with rate $r$,

$$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$

with effective receptive field

$k_{eff} = k + (k-1)(r-1).$

It then specializes to anisotropic strip kernels. For a vertical strip $(k \times 1)$,

$y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$

and for a horizontal strip $(1 \times k)$,

$y(u,v) = \sum_{j} w(0,j)\, x(u, v + r\,j).$

The reported design uses only horizontal and vertical strip convolutions; no diagonal strips are reported (Liu et al., 17 Jul 2025).

The motivation is geometric. DLOs are thin, elongated, piecewise-linear structures, and anisotropic strip kernels aligned with horizontal and vertical axes can attend efficiently over long, thin patterns while suppressing isotropic clutter. With dilation, these kernels expand the receptive field without a proportional increase in parameters, which is especially useful for cables and wires viewed from oblique angles (Liu et al., 17 Jul 2025).

The computational advantage is stated explicitly. A standard $1 \times 1$0 convolution has parameter count $1 \times 1$1, whereas a $1 \times 1$2 or $1 \times 1$3 strip convolution has $1 \times 1$4. Replacing a $1 \times 1$5 kernel by two orthogonal strips $1 \times 1$6 yields approximately $1 \times 1$7 rather than $1 \times 1$8, which the paper illustrates with the case $1 \times 1$9: $y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$0 versus $y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$1, or 60% fewer parameters/FLOPs (Liu et al., 17 Jul 2025).

ASCSPP applies this principle at the multi-scale pyramid level. Classic ASPP uses parallel atrous $y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$2 convolutions with multiple dilation rates. ASCSPP replaces these with parallel atrous strip convolutions—horizontal and vertical branches at multiple dilation rates—to decompose $y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$3 processing into lightweight, orientation-sensitive operations. Parallel branches are concatenated or merged, and the paper states that original low-level features are added before output to compensate for spatial detail loss in deeper features. No attention weighting among branches is reported (Liu et al., 17 Jul 2025).

The paper positions ASCSPP against related pyramid modules such as SPASPP and DAPP, which stack serial atrous convolutions and feature re-stacking. In the reported setup, ASCSPP preserves a parallel structure and prioritizes low cost via strip plus atrous design. The stated empirical outcome is that ASCSPP achieves the highest mIoU among the tested pyramid modules with only a minor FPS drop relative to ASPP (Liu et al., 17 Jul 2025).

5. Training protocol, data regime, and evaluation criteria

The training and evaluation setup is tightly coupled to the deployment scenario in (Liu et al., 17 Jul 2025). ASCnet is trained with the Adam optimizer, learning rate $y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$4, batch size $y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$5, for 200 epochs, with the backbone initialized using pre-trained weights. The training environment is Ubuntu 20.04, Python 3.8, and an NVIDIA RTX 3070 with 8 GB memory (Liu et al., 17 Jul 2025). The paper does not explicitly state the loss function; cross-entropy, Dice, focal loss, balancing, and label smoothing are all unspecified.

The reported lightweight models, including ASCnet, are trained and inferred at $y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$6. RT-DLO is additionally evaluated at $y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$7, described as its native resolution (Liu et al., 17 Jul 2025). For domain augmentation on the custom dataset, the authors used SAM to cut out DLOs and replaced backgrounds with 10 different backgrounds. No other augmentations, such as flip, rotate, or jitter, are reported (Liu et al., 17 Jul 2025).

The training data consist of two sources. The auto-generated wires dataset of Zanella et al. (2021) contributes 7,200 RGB images at $y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$8, partitioned into 5,760 training and 1,440 validation images. A self-built dataset contributes 1,850 RGB images at $y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u+i, v+j),$9, partitioned into 1,480 training and 370 validation images, and includes additional white and black DLOs. Across both sources, the total number of training images is 9,050, and all training images are described as top-down viewpoint data (Liu et al., 17 Jul 2025).

Testing is conducted on a self-built set of 548 real-world RGB images at $r$0, captured by a RealSense D455. Two scenarios are specified. Scenario 1 contains 391 images with varied floor colors—yellow, dark gray, and original gray-white—and DLOs in black, white, and yellow. Scenario 2 contains 148 images with similar setup but additional distractors such as baseboards that introduce line-like false positives. The stated purpose is to test robustness to viewpoint changes and background similarity (Liu et al., 17 Jul 2025).

The evaluation metrics include per-class IoU,

$r$1

mean IoU,

$r$2

as well as Precision $r$3, Recall $r$4, and Dice/F1. The task is binary segmentation, DLO versus background (Liu et al., 17 Jul 2025).

6. Sliding Window post-processing and quantitative performance

The Sliding Window (SW) module is a post-processing stage designed to exploit temporal consistency. The paper argues that lightweight models trained at low resolution, such as $r$5, often produce flickering false positives in real scenes with oblique viewpoints and clutter. SW assumes that true DLO regions persist across frames, whereas false positives appear intermittently (Liu et al., 17 Jul 2025).

Its algorithm is given in the paper as follows. The inputs are a binary mask $r$6, erosion kernel size $r$7, and sliding window length $r$8. First, morphological erosion is applied, $r$9, with the paper using $$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$0. Second, contours are extracted and small ones with area $$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$1 are ignored; the centroid of each connected component is computed. Third, IDs are assigned by matching current centroids to previous-frame centroids using Euclidean distance: if the distance is less than 50, the tracked ID is preserved; otherwise a new ID is assigned. Fourth, a queue of IDs over the last $$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$2 frames is maintained and frequencies are counted for sliding-window voting (Liu et al., 17 Jul 2025).

The headline system-level result is that ASC-SW achieves mIoU = 75.3% on the self-built dataset and 9.3 FPS on the Jetson Orin Nano edge device (Liu et al., 17 Jul 2025). The paper also reports raw segmentation results before SW and improved results after SW.

Before SW, ASCnet is the second-best model in both scenarios among the compared methods. In Scenario 1 at $$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$3, ASCnet reports $$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$4, $$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$5, $$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$6, and $$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$7, while RT-DLO at $$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$8 has the highest pre-SW mIoU of $$y(u,v) = \sum_{i} \sum_{j} w(i,j)\, x(u + r\,i, v + r\,j),$$9 (Liu et al., 17 Jul 2025). In Scenario 2, ASCnet reports $k_{eff} = k + (k-1)(r-1).$0, $k_{eff} = k + (k-1)(r-1).$1, $k_{eff} = k + (k-1)(r-1).$2, and $k_{eff} = k + (k-1)(r-1).$3, again ranking second before SW (Liu et al., 17 Jul 2025).

After SW, the ranking changes. In Scenario 1, DeepLabv3+-ResNet101 attains the highest mIoU at $k_{eff} = k + (k-1)(r-1).$4, while ASCnet reaches $k_{eff} = k + (k-1)(r-1).$5, $k_{eff} = k + (k-1)(r-1).$6, $k_{eff} = k + (k-1)(r-1).$7, and $k_{eff} = k + (k-1)(r-1).$8, described as within approximately 1% mIoU of the best result (Liu et al., 17 Jul 2025). In Scenario 2, ASCnet becomes the best-performing model after SW with $k_{eff} = k + (k-1)(r-1).$9, $(k \times 1)$0, $(k \times 1)$1, and $(k \times 1)$2 (Liu et al., 17 Jul 2025).

The improvement attributable to SW is large for ASCnet: Scenario 1 mIoU rises from $(k \times 1)$3 to $(k \times 1)$4, an increase of 11.7 points, and Scenario 2 rises from $(k \times 1)$5 to $(k \times 1)$6, an increase of 23.6 points (Liu et al., 17 Jul 2025). This suggests that temporal denoising is not an ancillary refinement but a major contributor to the practical performance of the overall ASC-SW system in cluttered, oblique-view environments.

Ablation studies further isolate the contribution of ASCnet’s internal modules. With channel attention, Scenario 1 mIoU increases from $(k \times 1)$7 to $(k \times 1)$8 and Scenario 2 from $(k \times 1)$9 to $y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$0 (Liu et al., 17 Jul 2025). Among multi-scale heads under the same framework and training strategy, the reported values are ASPP: mIoU $y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$1, FPS $y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$2; DAPP: mIoU $y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$3, FPS $y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$4; SPASPP: mIoU $y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$5, FPS $y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$6; and ASCSPP: mIoU $y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$7, FPS $y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$8. In that ablation, ASCSPP has the best accuracy with a small FPS drop from ASPP (Liu et al., 17 Jul 2025).

7. Relation to adjacent methods and limitations of the reported formulation

ASCnet is explicitly formulated as a lightweight alternative to heavier DLO detection models and large backbone segmentation systems. Relative to DeepLabV3Plus-MobileNetV2, it keeps nearly identical parameter count and GFLOPs while replacing standard ASPP with a DLO-oriented pyramid head and adding lightweight channel attention (Liu et al., 17 Jul 2025). Relative to DeepLabV3Plus-ResNet101 and RT-DLO at $y(u,v) = \sum_{i} w(i,0)\, x(u + r\,i, v),$9, it operates in a substantially lower-compute regime while still producing competitive or superior post-processed results in the reported scenarios (Liu et al., 17 Jul 2025).

Its design also differs from more generic multi-scale atrous modules. The paper contrasts ASCSPP with SPASPP and DAPP, which rely on serial atrous convolution and feature re-stacking. ASCSPP preserves a parallel structure and uses horizontal and vertical strip kernels to target linear geometry directly (Liu et al., 17 Jul 2025). This indicates a task-specific prior: orientation-sensitive anisotropic filtering is treated not merely as an efficiency measure but as a structural bias aligned with DLO appearance.

Several limits of the specification are also clear. The exact upsampling layer types, strides, padding, and classifier head details of the decoder are not enumerated. The loss function is not reported. The specific kernel sizes and dilation rate values used in the strip branches are not enumerated in the text. The paper also does not state whether ASCSPP’s strip convolutions are depthwise separable, even though the overall model remains within roughly 1% of DeepLabV3Plus-MobileNetV2 in parameters and FLOPs (Liu et al., 17 Jul 2025).

A further practical limitation emerges from the dataset design. All 9,050 training images are top-down viewpoint images, while testing focuses on oblique-view real-world images collected from a robot platform (Liu et al., 17 Jul 2025). The strong effect of SW post-processing suggests that temporal consistency compensates in part for viewpoint-domain mismatch and low-resolution inference. A plausible implication is that the segmentation network alone, without temporal filtering, remains vulnerable to distractors and unstable detections in deployment scenes containing line-like structures such as baseboards.

Within the published scope, ASCnet is therefore best understood as a specialized segmentation backbone for DLO-aware robotic navigation rather than as a general-purpose line-segmentation architecture. Its contribution lies in combining MobileNetV2 efficiency, DeepLabv3+-style fusion, and atrous strip-based multi-scale encoding with a temporal denoising stage that materially improves robustness on edge hardware (Liu et al., 17 Jul 2025).