Deep learning assisted particle identification in photoelastic images of granular flows

The transmission of forces within high speed granular flows may be straightforwardly viewed in two-dimensional photoelastic experiments, but precise measurements have remained elusive due to difficulties in differentiating between particles and forces with sufficient accuracy at reasonable processing speeds. This paper presents a novel approach to detect the positions of disks embedded in this complex situation, which is a crucial step in applying the methodologies necessary for the analysis of the photoelastic response of individual disks. We have developed a Deep Learning based solution to perform the segmentation of experimental photoelastic images, disentangling with high fidelity the disk outlines from the rest of each image. The accuracy and the reliability of the proposed methodology are discussed in detail, demonstrating that this approach can be effectively adopted for the problem under investigation, improving the quality of the photoelastic analysis and dramatically accelerating the data processing procedure.