Creation and segmentation of synthetic images of granular materials

ABSTRACT: Software packages based on image analysis techniques may be used to determine the particle size distribution (PSD) of coarse materials in the field based on photographs. Image segmentation is currently the main approach available to determine the PSD. Segmentation is used to delimit each particle to distinguish the superimposed particles from the isolated particles and the fine areas. The earliest approaches for image segmentation relied on applying Gaussian filters to an image, while more recent techniques use deep neural networks. A major bottleneck in neural network-based approaches is the availability of training data. This article proposes to create synthetic training photographs of granular materials and their corresponding segmented images, without the need for any manual adjustment. We used the Unity game engine to generate a large number of hyper-realistic training images, including their associated PSDs, segmentation, and photographic depth of field, given parameters such as the desired PSD and the sample weight.