DSWMamba: A deep feature fusion mamba network for detection of asphalt pavement distress

Abstract

The maintenance of asphalt pavements is vital for national infrastructure. Timely detection of pavement distress ensures optimal road performance and extends pavement lifespan. However, the various distress types and complex detection environments pose significant challenges. To overcome these, we propose a novel framework named DSWMamba, which is the first application of Mamba in asphalt pavement distress detection. This model incorporates targeted optimizations via the Selective Scanning Modules (SSM) to meet the unique demands of distress detection. To overcome the limitations of the SSM in receptive fields and poor image localization, we develop the Depth Fusion Selective Scan Block (DFSS), which integrates a Fusion Perception Block (FP-Block). This design expands the receptive field and strengthens local feature extraction. Additionally, the Vision Shuffle Down module(VSD) preserves both multiscale feature information, further enhancing feature representation. Unlike Transformer-based models, DSWMamba leverages a Separate Multidimensional Attention (SMD-Attention) mechanism for multi-path fusion, enabling global modeling. To improve the detection of large-scale features, we develop a Dynamic Separable align Head (DSA-Head), which decouples classification and localization tasks, significantly boosting feature recognition accuracy. Extensive experiments demonstrate that DSWMamba surpasses 18 target detection models, including CNN, Transformer frameworks and the latest YOLO and RTDETR series, on three datasets. DSWMamba improves transverse crack detection accuracy on the Asphalt Pavement Distress Dataset by 13.1% compared to Mamba YOLO while reducing parameters by 34.6% and computational cost by 44.6%. We release the Asphalt Pavement Distress Dataset, comprising 13,129 images of six distress types, offering a valuable resource for future research.