Gravelly soil uniformity identification based on the optimized Mask R-CNN model

Abstract

The uniformity of gravelly soil has an important influence on compaction quality. The most important task to judge the uniformity of gravelly soil is to segment the gravels from the image. However, gravels are widely and densely distributed, and their particle size varies greatly, increasing segmentation difficulty. Among existing studies, research on rapid and quantitative judgment methods of gravelly soil uniformity remains scarce. To address the abovementioned issue, a gravelly soil uniformity identification based on the optimized Mask R-CNN model is proposed. The original Mask R-CNN only produces one combined mask of multiple overlapping gravels, which hinders postprocessing and uniformity calculation. To address this problem, separate masks for each gravel are generated for better parameter calculation. Then, according to the characteristics of the pixel image of a single mask, the calculation of static moment is deduced and simplified. Finally, the single mask dataset of the optimized Mask R-CNN and static distance theory are used to establish a quantitative evaluation index of gravelly soil uniformity, in which the uniformity coefficient (UC) and area ratio coefficient (ARC) are adopted. In addition, the convergence curves and the Average Precision (AP) of the ResNet101 and the ResNet50 backbones are compared, and the result proves the superiority of ResNet101 in gravel segmentation. Furthermore, three data enhancement methods (namely, rotation, mirroring, and brightness transformation) are adopted to improve the AP performance and result in a 2.32% increase. The application in a real large-scale hydropower project shows that the AP can reach 88.96%, and each calculation and analysis can be controlled within one minute, which shows the effectiveness, convenience and efficiency of the method.