Hybrid graph convolutional and deep convolutional networks for enhanced pavement crack detection

Pavement crack detection plays a crucial role in ensuring traffic safety and prolonging service life. Cracks often exhibit irregular curved contours, with their local details submerged within complex background textures. Accurate segmentation of crack contours necessitates a model capable of capturing both long-range correlations and local details. While various methods have been proposed for complex crack detection challenges, most methods in the literature often encounter difficulties when dealing with the dual demands of global correlations and local details. This paper proposes a hybrid model that integrates graph convolutional networks (GCNs) for modeling long-range correlations and deep convolutional networks (DCNs) for capturing local textural details. The proposed model first transforms crack images into graphs using superpixel segmentation, where graph nodes and edges are initialized based on superpixels and their spatial adjacency. A hybrid deep learning architecture, comprising a GCN and an encoder-decoder DCN, is then designed. The GCN learns hidden representations of graph nodes through topology-adaptive convolution and edge message passing, capturing global topological features and long-range spatial dependencies. These features are complementarily enhanced with local spatial features extracted by the encoder-decoder DCN. Experimental results on benchmark pavement crack detection datasets demonstrate that the proposed method achieves competitive detection performance with fewer learnable parameters, accurately capturing both the slender, irregular crack contours and their local textural details. Thus, the method offers a reliable solution for pavement crack detection with improved generalization and interpretability.