Abstract: To address the problems that the traditional rock chips screening test is time-consuming and laborious, which is difficult to provide real-time guidance for construction, and the dataset of the existing rock chips image segmentation models is single in content, this study carries out research on high-precision segmentation model of rock chips images and prediction of rock chips particle size distribution. Relying on three typical shield tunnel construction projects, rock chips samples are collected from tunnel boring sections, and rock chips images of different lithologies are obtained by using a self-built indoor shooting platform to construct an image dataset with sufficient data volume and various types of rock chips. The Swin-Unet deep learning network is adopted to train the rock chips image dataset to achieve high-precision segmentation, and its performance is compared with other mainstream models on the same dataset. Post-processing is performed on the segmented rock chips images to extract the corresponding two-dimensional image parameters. The particle size distribution law of rock chips is predicted based on the principle of area method, and the predicted results are compared with the rock chips screening curve to verify its rationality. The results show that under the same training period, the segmentation indexes of ACC, Dice, IoU and HD95 of the Swin-Unet model are 89.51%, 85.67%, 75.65% and 11.91 mm, respectively, which are better than those of the comparison algorithms, with relatively small parameters and computational cost. The predicted results of rock chips particle size distribution are generally consistent with the rock chips screening curve. The predicted values are slightly lower than the screening values when the rock chips particle size is less than 16 mm, and the opposite is true when the particle size is larger than 16 mm. The error of the rock chips roughness index calculated based on the prediction results of the Swin-Unet model is within 5%, indicating a satisfactory prediction effect.