Spectrally Filtered Optical Convolutional Neural Network for Shift-Robust Bright-Field Classification

Abstract

Optical neural networks (ONNs) offer high-throughput and energy-efficient computation, providing a promising alternative to conventional electronic systems. However, their application in bright-field microscopy is limited by random spatial displacements of dynamic targets and a strong unmodulated transmitted-light component concentrated at the zero order of the spatial-frequency domain. This direct current (DC) component significantly reduces feature contrast, posing a major challenge for reliable optical inference. To address these challenges, we propose a spectrally filtered optical convolutional neural network (SF-OCNN) that integrates Fourier filtering, optical convolution, and an expanded detection region. The Fourier filter suppresses the DC component before optical convolution while preserving high-frequency structural information. Optical convolution provides inherent translation equivariance, while the expanded detection region accommodates the displacement of the output response during inference. The proposed system is validated through both numerical simulations and physical experiments. It achieves validation accuracies of 91.12% in simulation and 88.50% in the physical experiment on a bright-field MNIST data set (digits 0–3). For experimentally acquired microfluidic cell images re-encoded on the DMD, the corresponding accuracies are 99.50% and 98.25%. These results demonstrate that combining physical DC suppression with optical convolution provides an effective approach for robust optical computing in realistic bright-field environments.