Lightweight Yet Effective: A Modular Approach to Crack Segmentation

Abstract

Automatic crack detection is criticalfor road safety. However, existing models face challenges due to complicated cracks, difficult backgrounds, and computational inefficiency. This impedes real-world applicability, especially on mobile platforms. To address these limitations, we propose a lightweight yet robust crack segmentation model based on a modular architecture. It comprises four main modules: Parallel Feature Module (PFM) for multi-feature extraction, Edge Extraction Module (EEM) to obtain the outer shape of the cracks, Pixel-wise Dilation and Attention Module (PDAM) applying pixel-wise attention, Feature Reduction and Concatenation Module (FRCM) for efficient feature fusion. The proposed model incorporates conventional image processing methods within the CNN framework to balance efficiency and performance. Evaluated on Crack500, DeepCrack, GAPs384, AigleRN-TRIMM, and ShadowCrack datasets, the proposed model achieves state-of-the-art performance among existing lightweight models on multiple metrics, while requiring only