Treffer: Application of U-Net+ + deep learning network for segmentation and processing of asphalt mixture X-ray CT images.

The robustness and generalization ability of traditional image segmentation algorithms are poor, and their performance is further degraded by factors such as uneven lighting, photoelectric noise, and low contrast, making it challenging to segment three-phase materials in X-ray CT images of asphalt mixtures. With the advancement of computer network technology, deep learning image segmentation algorithms have garnered widespread attention and application. However, the specific implementation process of deep learning network algorithms for X-ray CT images of asphalt mixtures remains unclear, and there is a notable absence of publicly available datasets specific to asphalt mixtures. Therefore, this study explores the specific implementation process of the U-Net+ + algorithm in X-ray CT images of asphalt mixtures. Deep learning networks were constructed using Python and PyTorch, implementing U-Net, U-Net+ + networks, and AG-U-Net and AG-U-Net+ + networks, each incorporating the AG attention mechanism. A specific method for creating a deep learning dataset was proposed using threshold segmentation and manual annotation of a custom-labeled dataset. Three-phase material segmentation and 3D modeling of OGFC-16 and AC-16 Marshall specimens were successfully achieved through neural networks. The research results indicate that image cropping during the dataset production process may result in the loss of global information, leading to segmentation errors for porous targets. In the segmentation of aggregates and pores, AG-U-Net+ + and U-Net+ + exhibit different optimal evaluation metrics, highlighting the need for further optimization of the dataset creation scheme for pore targets in subsequent research. • Implemented the segmentation of asphalt mixture X-ray CT images using U-Net+ + network and AG-U-Net+ + models. • Proposed an effective deep learning dataset processing method for X-Ray CT images of asphalt mixtures. • Evaluated the image segmentation performance of asphalt mixtures based on the confusion matrix. • Reconstructed the 3D model of asphalt mixture, providing a foundation for further analysis. [ABSTRACT FROM AUTHOR]