Synthetic Image Generation for Visual Particle Size Distribution Estimation Based on U-NET Convolutional Networks

Many applications in the field of bulk material handling require the determination of the Particle Size Distribution (PSD) of the handled materials. Typically, this can only be achieved by manual sampling and a subsequent laboratory analysis. However, the availability of novel image segmentation algorithms based on Convolutional Neural Networks (CNNs) leads to the possibility for an implementation of visual PSD estimation techniques. One prominent example of such approaches is the U-Net image segmentation architecture. However, one fundamental problem for the training of the underlying CNN is the difficulty to gather ground truth data, since the exact correspondence between a set of images and the underlying particle sizes cannot be achieved in practice. This paper introduces a novel synthetic image generation procedure for the visual PSD estimation by using the U-NET structure. The overall approach is validated and verified by using a prototypic implementation of a U-NET segmentation pipeline. It is shown, that the proposed synthetic image dataset generation can be applied for all visual PSD estimation based on machine learning algorithms.