An investigation into the application of neural networks for optical vertex image segmentation

This study investigates the possibility of segmenting modeled images of optical vortices using neural networks. An original dataset was collected for the training and testing of neural networks. The dataset consists of 1000 images of optical vortices passing through circular gratings with varying heights. These images were obtained using the finite-difference time-domain (FDTD) method. Experiments evaluated optimal neural network configurations, including architectures like U-Net, U-Net++, and variations with different encoders based on the ResNet, as well as loss functions like cross-entropy, focus, and dice. The best configuration used the original U-Net architecture with cross-entropy loss, achieving an average dice metric score of 0.87. The results of this study demonstrate the feasibility of applying neural networks for image segmentation of radiation propagation on features with different elevations of the topography. This study contributes to the growing amount of work on the integration of neural networks into optical systems and emphasizes the advantages of using these advanced algorithms to solve complex problems in the field.