Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Prototype Learning-Based Few-Shot Segmentation for Low-Light Crack on Concrete Structures

Yulun Guo

TL;DR

The paper tackles crack segmentation under challenging low-light conditions with limited labeled data. It introduces a dual-branch prototype network guided by Retinex theory, augmented by a Cross-Similarity Prior Mask Generation module and a Multi-Scale Feature Enhancement module to enable robust few-shot segmentation. Key contributions include end-to-end training for low-light crack segmentation, explicit cross-image priors for crack localization, and attention-driven multi-scale feature fusion, validated on real and synthetic datasets with state-of-the-art results. The approach reduces annotation burden while achieving reliable performance in difficult lighting, and code is publicly available for reproducibility.

Abstract

Crack detection is critical for concrete infrastructure safety, but real-world cracks often appear in low-light environments like tunnels and bridge undersides, degrading computer vision segmentation accuracy. Pixel-level annotation of low-light crack images is extremely time-consuming, yet most deep learning methods require large, well-illuminated datasets. We propose a dual-branch prototype learning network integrating Retinex theory with few-shot learning for low-light crack segmentation. Retinex-based reflectance components guide illumination-invariant global representation learning, while metric learning reduces dependence on large annotated datasets. We introduce a cross-similarity prior mask generation module that computes high-dimensional similarities between query and support features to capture crack location and structure, and a multi-scale feature enhancement module that fuses multi-scale features with the prior mask to alleviate spatial inconsistency. Extensive experiments on multiple benchmarks demonstrate consistent state-of-the-art performance under low-light conditions. Code: https://github.com/YulunGuo/CrackFSS.

Prototype Learning-Based Few-Shot Segmentation for Low-Light Crack on Concrete Structures

TL;DR

The paper tackles crack segmentation under challenging low-light conditions with limited labeled data. It introduces a dual-branch prototype network guided by Retinex theory, augmented by a Cross-Similarity Prior Mask Generation module and a Multi-Scale Feature Enhancement module to enable robust few-shot segmentation. Key contributions include end-to-end training for low-light crack segmentation, explicit cross-image priors for crack localization, and attention-driven multi-scale feature fusion, validated on real and synthetic datasets with state-of-the-art results. The approach reduces annotation burden while achieving reliable performance in difficult lighting, and code is publicly available for reproducibility.

Abstract

Crack detection is critical for concrete infrastructure safety, but real-world cracks often appear in low-light environments like tunnels and bridge undersides, degrading computer vision segmentation accuracy. Pixel-level annotation of low-light crack images is extremely time-consuming, yet most deep learning methods require large, well-illuminated datasets. We propose a dual-branch prototype learning network integrating Retinex theory with few-shot learning for low-light crack segmentation. Retinex-based reflectance components guide illumination-invariant global representation learning, while metric learning reduces dependence on large annotated datasets. We introduce a cross-similarity prior mask generation module that computes high-dimensional similarities between query and support features to capture crack location and structure, and a multi-scale feature enhancement module that fuses multi-scale features with the prior mask to alleviate spatial inconsistency. Extensive experiments on multiple benchmarks demonstrate consistent state-of-the-art performance under low-light conditions. Code: https://github.com/YulunGuo/CrackFSS.
Paper Structure (17 sections, 21 equations, 6 figures, 2 tables)

This paper contains 17 sections, 21 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Low-light Crack Segmentation
  4. Few-Shot Segmentation
  5. Methodology
  6. Task Description
  7. The Overall Architecture
  8. Cross-Similarity Prior Mask Generation Module
  9. Multi-Scale Feature Enhancement Module
  10. Loss Function
  11. Experiments
  12. Datasets
  13. Implementation Details
  14. Quantitative Results
  15. Qualitative Results
  16. ...and 2 more sections

Figures (6)

  • Figure 1: Different strategies for low-light crack segmentation: (a) Directly applying the same segmentation model to images captured under varying illumination conditions results in inconsistent segmentation performance; (b) Employing different image enhancement models introduces additional uncertainty in the segmentation results produced by the same segmentation model.
  • Figure 2: Architecture overview of the proposed network. The framework consists of five parts: (a) dual feature extractors processing RGB images and corresponding reflectance component; (b) a Cross-Similarity Prior Mask Generation (CSPMG) module hat produces query prior masks from high-level feature; (c) a Multi-Scale Feature Enhancement (MSFE) module that refines query features through support-guided attention; (d) a Prototype Fusion Module (PFM) that adaptively updates support prototype; (e) a Self-Support Prototype (SSP) module that further updates the support prototype.
  • Figure 3: (a) The architecture of the Cross-Similarity Prior Mask Generation (CSPMG) module; (b) The architecture of the Multi-Scale Feature Enhancement (MSFE) module.
  • Figure 4: Prior mask generation and qualitative results on the LCSD dataset. Top: Support images with crack regions masked. Middle: Query images and corresponding enhanced visualizations. Bottom: Generated prior masks highlighting regions of interest in the query images.
  • Figure 5: Qualitative comparison of segmentation results on the LCSD dataset under the 1-shot setting. From left to right: (a) query image (contrast-enhanced for visualization), (b) ground truth, (c) MLC MLC, (d) SSP SSP, (e) CrackNex CrackNex, and (f) our model.
  • ...and 1 more figures