DcsNet: a real-time deep network for crack segmentation

Detecting cracks are a great significance for the maintenance of the man-made buildings, and deep learning methods such as semantic segmentation have greatly boosted this process in recent years. However, the existing crack segmentation methods often sacrifice feature resolution to achieve real-time inference speed which leads to poor performance, or use complex network module to improve the accuracy which leads to lower inference speed. In this paper, we propose a novel Deep Crack Segmentation Network (DcsNet) that incorporates two feature extraction branches to achieve the balance of speed and accuracy. We first design a morphology branch (MB) to preserve the morphology information of scale invariance that consists of a lightweight convolution network, a pyramid pooling module (PPM), and an attention module (CSA). Meanwhile, a shallow detail branch (DB) with a small stride is constructed to supplement detailed information. Extensive experiments are conducted on five challenging datasets (Crack500, Deepcrack, Gaps384, Structure, and Damcrack), and the results demonstrated that the proposed network achieves a good trade-off between accuracy and inference speed and outperforms state-of-the-art methods.

Automated summary

In this paper, a novel Deep Crack Segmentation Network (DcsNet) is proposed, which achieves a balance of speed and accuracy through two feature extraction branches. Extensive experiments show that the proposed network achieves a good trade-off between accuracy and inference speed, outperforming state-of-the-art methods.