CT Scan Image Segmentation of Asphalt Mixture Based on Improved U-Net

Abstract

In the CT scan image of asphalt mixture, there are common factors such as dense mixture area and uneven illumination, which result in low accuracy of local feature segmentation. Through the introduction of the attention mechanism in U-Net, before fusing the features of each resolution in the encoder with the relating features in the decoder, an attention mechanism is added to make the encoder readjust its output features, which is similar to imitating human attention to achieve the effect of paying attention to multiple details at the same time. Achieve channel enhancement of the local characteristic area of the asphalt mixture, and improving the segmentation ability of the network model to the local characteristic area. Those test outcomes indicate that, compared for the universal segmentation algorithm and the classic U-Net model segmentation algorithm, the methodology of segmentation in this paper has better performance in terms of MPA coefficient, MIoU coefficient and Dice coefficient.