Increased Computation Speed of Neural Network-Aided Computer Vision Via Coded Diffraction of Off-Axis Optical Vorticies

The current methods of image classification via deep-learning are slow and prohibitively computationally and energetically expensive. We establish a preprocessing technique that uses coded diffraction of off-axis optical vortices of varying topologies in the Fourier domain to improve the performance of quick Dense Neural Networks. The weights derived by a softmax-activated dense neural network are used to calculate the Fisher information matrix to a first approximation. We demonstrate the Green’s function for the Fisher Information contained within an image given a coded diffraction pattern. We also show that the signals with more Fisher Information are concentrated within the neural network to aid in the efficient classification of images. According to Muminov et al., on-axis vortices in dense neural networks are 200 times faster than traditional convolutional neural networks achieving similar performance. However, this off-axis technique improves on the accuracy of the on-axis case by 45%.