Rockfill material segmentation and gradation calculation based on deep learning

Abstract

The gradation of rockfill materials affected the rolling quality and body strength of earth-rockfill dam. To address the limitations of manual screening and traditional computer vision methods, a gradation detection method suitable for rockfill dam construction sites was proposed based on deep learning, which can output the detection results end-to-end. The method utilized the modified MASK R-CNN model to locate and segment rockfill materials. Specifically, the model used ResNeXt101 as the backbone network and adds the squeeze-and-excitation block to enhance the feature extraction capability of the model. The detection results of 500 test images showed that proposed method had high accuracy, the average precision and intersection over union value of the model reach 0.934 and 0.879. The morphological characteristics of rockfill particles were extracted by rotating calipers algorithm and used for grading calculation. Comparing this calculation method with manual screening, the results indicated that the overall fitting effect of the gradation detected by proposed method was in accord with the actual gradation parameters. The gradation of rockfill materials can be obtained within 60 s. This paper provided a new method for rapid analysis of particle size distribution, which can be used for auxiliary construction.