• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Sunday, November 22, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

CrackLIFT Integrates Learnable Wavelet Lifting with State-Space Modeling for Infrastructure Crack Segmentation

by
November 21, 2026
in Technology
Reading Time: 5 mins read
0
CrackLIFT Integrates Learnable Wavelet Lifting with State-Space Modeling for Infrastructure Crack Segmentation

CrackLIFT Integrates Learnable Wavelet Lifting with State-Space Modeling for Infrastructure Crack Segmentation

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Structural health monitoring across civil infrastructure such as bridges, dams, tunnels, and industrial facilities depends on identifying early indicators of structural distress, among which surface cracks represent the most prevalent and visible warning signs. Automated pixel-level crack segmentation offers a scalable alternative to labor-intensive visual inspections and can locate surface flaws in hazardous or inaccessible environments. However, automated crack detection remains computationally difficult due to the intrinsic physical characteristics of crack formations, which are typically thin, elongated, and geometrically irregular, occupying only a small fraction of the total image area while exhibiting weak visual contrast against surface substrates. Furthermore, real-world inspection images are frequently corrupted by high-frequency background noise, including water stains, surface scratches, uneven shadows, and concrete blemishes, that closely mimic true crack boundaries when analyzed through conventional spatial filters.

To overcome these persistent imaging obstacles, researchers have introduced CrackLIFT, a deep learning network designed specifically for crack segmentation that integrates learnable lifting-based wavelet decomposition directly with visual state-space modeling. Standard convolutional neural networks excel at extracting local edge cues but frequently fragment extended crack paths because their effective receptive fields cannot readily bridge long-range context without excessive downsampling. Conversely, vision transformers capture global spatial dependencies through self-attention mechanisms, yet their patch tokenization process can degrade boundary localization and blur thin fracture lines. State-space models, notably the selective scanning framework known as Mamba, achieve long-range spatial context modeling while maintaining linear computational scaling with sequence length. CrackLIFT builds upon these selective scanning capabilities by integrating an adaptive frequency decomposition inside each encoder stage, ensuring that structural context and fine edge responses are simultaneously preserved and coordinated.

A fundamental limitation in prior frequency-aware neural architectures has been the reliance on static mathematical transformations, such as the fixed discrete wavelet transform, which imposes a predetermined sub-band frequency partition that cannot adapt to the segmentation loss during training. In contrast, CrackLIFT is constructed around specialized lifting-based interactive frequency transformation blocks that parameterize the traditional predict and update steps of a wavelet lifting scheme. In each block, the learnable adaptive wavelet decomposition module splits the two-dimensional feature tensor along horizontal and vertical spatial dimensions into even- and odd-indexed samples. Lightweight neural networks consisting of depthwise separable convolutions, Gaussian Error Linear Unit activations, and point-wise convolutions parameterize the predict and update operators, allowing the frequency decomposition boundaries to be learned dynamically alongside the network parameters to optimize the crack segmentation objective.

The forward decomposition produces four distinct spatial frequency sub-bands: a low-frequency approximation that captures smooth structural background geometry, and three directional detail sub-bands representing horizontal, vertical, and diagonal high-frequency variations. To prevent excessive memory growth and parameter bloat, the three directional detail sub-bands are concatenated along the channel dimension and compressed via a one-by-one convolution into a unified high-frequency feature representation matching the stage channel width. This unified high-frequency stream and the corresponding low-frequency approximation are then routed into dual, frequency-specialized state-space processing streams. Within these parallel paths, independent two-dimensional selective scan operators scan the feature grids along both width and height axes, capturing long-range spatial dependencies while retaining linear computational complexity relative to the spatial resolution.

Because isolated processing streams risk neglecting the cross-scale dependencies that tie thin crack boundaries to broader structural contexts, each transformation block incorporates a bidirectional frequency interaction module. This module enables simultaneous cross-stream feature modulation without introducing serial computational bottlenecks. In the low-to-high frequency direction, the model derives both a spatial attention gate via a point-wise convolution and a channel attention gate via global average pooling and a multi-layer perceptron from the pre-processed low-frequency state; multiplying these gates against the high-frequency stream effectively suppresses extraneous background noise in non-crack regions. Simultaneously, in the high-to-low direction, the module extracts an edge attention map from the high-frequency features through depthwise and point-wise convolutions, using a learnable scaling parameter initialized at 0.1 to inject subtle boundary cues back into the low-frequency stream without disrupting foundational contextual features.

Following the bidirectional exchange, the features undergo wavelet-consistent reconstruction to return to the original spatial resolution of the block. A spatial split head composed of depthwise separable and point-wise convolutions regenerates the three distinct directional high-frequency sub-bands from the unified high-frequency stream. These regenerated sub-bands and the processed low-frequency representation are passed through an exact inverse lifting transformation, which reverses the forward mathematical operations using the same shared predict and update networks while inverting their signs. The resulting reconstructed feature map is added back to the original block input through a residual connection, preserving stable gradient propagation throughout the four hierarchical stages of the encoder backbone, which feature channel capacities of 32, 64, 128, and 256 across uniform depths of two blocks per stage.

To guide the intermediate representations during training, the network optimizes a compound loss function that combines standard pixel-wise cross-entropy on the final segmentation mask with an auxiliary wavelet-guided frequency alignment loss, weighted by a factor of 0.5. The alignment loss constructs stage-matched frequency reference targets by resizing the ground-truth binary mask and applying a fixed Haar discrete wavelet transform to yield exact low-frequency and directional detail targets. Auxiliary training-only convolutional projection heads map the intermediate low- and high-frequency encoder features to these targets to calculate an L1 sub-band alignment loss, balanced by weight factors of 0.2 for low frequency and 1.0 for high frequency. Additionally, an inverse Haar reconstruction consistency term compares the reconstructed mask proxy directly against the resized ground truth, ensuring that frequency representations retain physical fidelity without adding computational overhead during inference.

Empirical evaluation on the comprehensive CrackSeg9k benchmark demonstrated that CrackLIFT attained superior segmentation performance relative to thirteen competing convolutional, transformer, and state-space architectures. The model achieved a mean Intersection over Union of 82.10 percent, an F1-score of 88.75 percent, and a Boundary F1-score of 84.48 percent using a standard two-pixel distance tolerance. When evaluated against U-Mamba, the strongest competing state-space baseline, CrackLIFT improved mean Intersection over Union by 0.75 percentage points and Boundary F1-score by 0.93 percentage points. Furthermore, CrackLIFT demonstrated notable computational efficiency, requiring 9.70 million parameters and 4.77 GFLOPs for 400-by-400 input images, which represents 47.6 percent fewer parameters, 68.6 percent fewer GFLOPs, and a 41.7 percent higher inference throughput than U-Mamba.

The statistical validity of these performance gains was confirmed through repeated-run evaluations on CrackSeg9k across five independent random seeds. Across all five trials, CrackLIFT maintained its performance advantage, outperforming U-Mamba by 0.75 points in mean Intersection over Union, 0.74 points in F1-score, and 0.95 points in Boundary F1-score, while exceeding the SCSegamba baseline by 1.19, 0.87, and 1.31 points across the respective metrics. Two-sided paired t-tests confirmed that all comparative performance improvements were statistically significant with p-values of 0.007 or lower. Evaluations across boundary distance tolerances ranging from one to five pixels further demonstrated that CrackLIFT maintained top performance across all scales, exceeding the second-best method by 1.55 Boundary F1 points under the most rigorous one-pixel threshold.

The framework also exhibited consistent generalization across material-specific datasets, achieving 81.88 percent mean Intersection over Union and 84.62 percent Boundary F1-score on the Asphalt3k dataset, alongside 87.05 percent mean Intersection over Union and 89.55 percent Boundary F1-score on Concrete3k, outperforming all evaluated baselines on both surfaces. In synthetic degradation benchmarks evaluating resilience to additive Gaussian noise, linear motion blur, and global brightness variations across three severity tiers, CrackLIFT retained higher accuracy than competing models, widening its margin over U-Mamba as perturbation severity escalated. Controlled ablation studies validated that learnable lifting decomposition, bidirectional interaction, and auxiliary frequency alignment each contributed measurable accuracy gains, although failure analyses revealed that elongated non-crack artifacts like expansion joints and dense branching networks still pose occasional disambiguation challenges.

Subject of Research: Deep learning architectures combining learnable wavelet transforms and state-space models for automated structural crack segmentation

Article Title: CrackLIFT: Learnable lifting-based wavelet decomposition with state-space modeling for crack segmentation

Article References: Li, Y., Ma, Z., Zhang, Q., & Xiao, L. (2026). CrackLIFT: Learnable lifting-based wavelet decomposition with state-space modeling for crack segmentation. Journal of King Saud University Computer and Information Sciences, 38(9), Article 970. https://doi.org/10.1007/s44443-026-01379-x

Image Credits: AI Generated

DOI: 10.1007/s44443-026-01379-x

Keywords: crack segmentation, CrackLIFT, state-space models, Mamba architecture, lifting scheme, wavelet decomposition, structural health monitoring, CrackSeg9k, Asphalt3k, Concrete3k, computer vision, image segmentation, frequency analysis, deep learning, infrastructure inspection, wavelet transform

News Source: Blake Davidson. (November 21, 2026). CrackLIFT Integrates Learnable Wavelet Lifting with State-Space Modeling for Infrastructure Crack Segmentation. Scienmag.

Tags: Asphalt3kComputer VisionConcrete3kcrack segmentationCrackLIFTCrackSeg9kdeep learningfrequency analysisimage segmentationinfrastructure inspectionlifting schemeMamba Architecturestate space modelsStructural Health Monitoringwavelet decompositionwavelet transform
Share12Tweet7Share2ShareShareShare1

Related Posts

Open Source QGIS Plugin Automates Multi Method Spatial Interpolation for Agriculture

Open Source QGIS Plugin Automates Multi Method Spatial Interpolation for Agriculture

November 22, 2026
Harmony Framework Unifies Modeling and Stability Analysis for Hybrid Power Grids

Harmony Framework Unifies Modeling and Stability Analysis for Hybrid Power Grids

November 22, 2026

Open Source Framework Automates Geant4 Radiation Detector Prototyping and Code Generation

November 22, 2026

Neutrophil Drug Delivery Systems Target Inflammation and Tumors in Preclinical Studies

November 22, 2026

POPULAR NEWS

  • Review Outlines Advanced Drying Methods and Artificial Intelligence for Kaffir Lime Leaves

    Review Outlines Advanced Drying Methods and Artificial Intelligence for Kaffir Lime Leaves

    29 shares
    Share 12 Tweet 7
  • Emirates Internal Medicine Conference Highlights Advances Across Critical Care and Pharmacotherapy

    29 shares
    Share 12 Tweet 7
  • Genomic Analysis Distinguishes Five Diverse Subtypes of Metabolic Steatotic Liver Disease

    29 shares
    Share 12 Tweet 7
  • Italian Citizens Favor Farm Artificial Intelligence While Expressing Notable Response Uncertainty

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Endothelial FBXW7 Degrades NOTCH1 to Silence SEMA3G and Curb Liver Fibrosis

Open Source QGIS Plugin Automates Multi Method Spatial Interpolation for Agriculture

Herbal Nanovesicles Suppress Colorectal Tumors and Boost Immunotherapy in Mouse Models

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.