A deeply supervised attention network for pavement crack segmentation from unmanned aerial vehicle images

Abstract

Abstract Crack segmentation is crucial for evaluating the health condition of pavement. However, the various crack sizes, class imbalance issues, and background interference bring challenges to accurate segmentation of pavement cracks. To overcome these challenges, a deeply supervised attention network for pavement crack segmentation from unmanned aerial vehicle images (DSA-Net) is proposed, which is based on an encoder-decoder architecture. Specifically, to extract multi-scale crack features, a novel multi-scale encoder module (EM) is designed by combining dilated convolution and residual structure. Then, a left-side path is designed on the left side of the EM to alleviate the influence of class imbalance on feature extraction and help recover small-sized crack information. Next, an attention module with high-dimensional features guiding low-dimensional features (AM-HGL) is proposed to focus on crack-relevant features and suppress interference information during the feature decoding process. Furthermore, a deep supervision module is introduced to generate more accurate segmentation results and improve the learning ability of the segmentation network. Finally, a weighted loss function based on binary cross-entropy and Dice is introduced to further solve the class imbalance issues. To verify the effectiveness of the proposed DSA-Net, comprehensive experiments are conducted on a self-made unmanned aerial vehicles pavement crack dataset and a public pavement crack dataset. The quantitative and qualitative experimental results show that the proposed method achieves a better balance in segmentation performance, efficiency, and deployment cost compared to other state-of-the-art methods, and can meet the needs of pavement crack segmentation in practical application scenarios.