Asphalt mixture image segmentation by RAN-UNet based on attention mechanism and multi-scale feature fusion

Asphalt mixtures exhibit dimensional complexity and morphological diversity, with blurred component boundaries hindering accurate mesostructure restoration. A convolutional neural network model RAN-UNet based on attention mechanisms and multi-scale feature fusion was proposed in this paper. The threshold segmentation method, U-Net and its improved models were compared to process CT images of asphalt mixtures, and the model performance and segmentation accuracy were evaluated for aggregates and voids within asphalt mixtures of various gradations. The applicability of methods to different gradation and components of the asphalt mixture were analyzed. The results showed that RAN-UNet has a better predictive ability and robustness than U-Net and its commonly used improved models. It can predict the morphology of target objects more accurately and restore the angularity of aggregates to a greater extent. This study provide an efficient and highly accurate method for the field of asphalt mixture image segmentation.