betaFeedback

WorksComputer Science

FDAM: A Frequency-Direction-Aware Multi-Task Framework for Pavement Crack Grid Localization

Abstract

Pavement crack grid localization remains challenging for thin and tiny cracks due to weak visual cues and background interference. Box detection often misses these cracks, while supervised semantic segmentation faces limited pixel-wise annotations and high labeling costs. In this paper, we propose a Frequency-Direction-Aware Multi-Task framework (FDAM) with box detection and grid classification branches. The core grid classification branch comprises Frequency-Domain Enhancement (FDE) and Directional Feature Refinement (DFR). FDE adaptively weights spectral bands and fuses reconstructed features with spatial features. DFR captures thin crack patterns through multi-scale strip convolutions, direction-sensitive modulation, and feature recalibration. Sparse positive cells and ambiguous local predictions motivate adaptive grid supervision. Adaptive Importance Reweighting emphasizes uncertain predictions and local label transitions, with focal supervision for all cells. Our contribution is to couple the box detection and grid classification branches through shared feature learning for pavement crack grid localization. We validate FDAM on our RoadDistress2026 dataset of thin and tiny cracks and further evaluate it on public datasets Crack500 and DeepCrack. The code is available at https://github.com/Lab-PANbin/FDAM

Paper details

vlnc:W_y8g5j7vn
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

ChatGPT

View statement
Looking for related papers…