Research on 3D Particle Field Reconstruction Method Based on Deep Learning

Abstract

Reconstructing 3D particle fields from single holograms is an important topic in the computer vision field. To increase network accuracy, we incorporate a channel attention mechanism based on the U-Net architecture in this research; meanwhile, we use the LeakyReLU activation function to accelerate the network convergence. Deep learning is used to extract information from a single hologram that can recreate the 3D particle field. The network receives the digital hologram as input, and the radius and 3D locations of the particles are converted into 2D grayscale images as real labels. Simulation and experimental results show that 2D grayscale images with clear edge textures can be rapidly encoded using the neural network, and the average SSIM and PSNR of the network output results with the real target on the test data set can reach 0.989 and 32.56.