Reconfigurable Time‐domain Convolutions for Imaging Processing With Synthetic Temporal Lattices

ABSTRACT Convolution operations are fundamental to signal processing and convolutional neural networks, forming the basis for feature extraction algorithms in image, speech, and data analysis. Optical systems offer ideal platforms for realizing convolution operations, overcoming electronic bottlenecks due to their inherent merits of parallelism and low energy consumption. However, spatial optical systems still suffer from fixed architectures and limited reconfigurability. Here, by harnessing a synthetic temporal lattice created in a fiber‐loop system, we propose and experimentally demonstrate time‐domain convolutions with reconfigurable convolution kernels. We map 1D signals onto a sequence of time‐encoded coherent optical pulses in the temporal lattice and implement convolution operations through predesigned intensity and phase modulations. By further applying a dimensional transformation algorithm, we also realize 2D convolution operations used for image processing, achieving typical functionalities of image blurring and edge recognition. The study paves a promising avenue for constructing flexible and reconfigurable convolution platforms using synthetic lattices, with potential applications for imaging processing, signal reshaping, and even neuromorphic computing.