SVGACrack: Sparse Vision Graph Attention Segmentation Networks Enabling Precise Pavement Crack Detection

Most current pavement crack detection tasks depend on Convolutional Neural Networks (CNNs) and Vision Transformer Networks (ViT) for feature extraction, and this dependency often makes it difficult to effectively capture irregular and complex cracks, which in turn affects crack detection performance. To tackle this challenge, this paper proposes a novel graph-based sparse attention mechanism encoder-decoder structure model for pavement crack detection (SVGACrack), which utilizes a hybrid CNN-GNN architecture MobileViG as the encoder backbone for multi-scale feature extraction in multiple stages, and the decoder adopts a lightweight multilayer perceptron structure, which uses an MLP layer first to unify the number of channels of the feature maps generated in each stage. The feature maps are then up-sampled and merged to capture both semantic abstraction and local details, improving detection accuracy. Finally, on the DeepCrack and Crack500 dataset, our proposed SVGACrack model obtains the highest F<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf>-score, which outperforms current the state-of-the-art methods and proves its effectiveness in pavement crack detection.