A particle shape extraction and evaluation method using a deep convolutional neural network and digital image processing

Abstract Acquiring particle shapes from raw particle images with complex backgrounds is a prerequisite for particle shape evaluation, yet it is a challenging task. In this study, a systematic framework for particle extraction and shape analysis is developed using a deep convolutional neural network (lightweight U-net) and digital image processing. First, raw images of particles are cropped and labeled manually to train the neural network. Then, the well-trained network is employed to extract particle projections from images of arbitrary size with complex backgrounds. Next, the particle boundaries are separated and smoothed using the improved erosion and flood filling method and the B-spline curve technique. Finally, the shapes of the extracted particles are evaluated and compared with the shape data obtained from the manually extracted particles. The shape distributions from these two approaches are found to be well correlated, illustrating the reliability and capability of the proposed algorithms.