All-Optical DCT Encoding and Information Compression Based on Diffraction Neural Network
He Ren, YuXiang Feng, Shuai Zhou, Di Wang, Xu Yang, Shouqian Chen
Abstract
As information technology advances and data volumes grow rapidly, there is an increasing requirement for information security and throughput in image transmission systems. Diffractive neural networks (DNNs), a novel all-optical information processing paradigm, offer several advantages including high-speed processing, low energy consumption, and high spatial utilization. These networks also leverage the powerful reverse-design capabilities of deep learning methods, enabling the efficient implementation of various image information encoding techniques. The discrete cosine transform (DCT), a well-established technology widely used in image encoding, shares features with DNNs, such as linear operations, spatial–frequency domain conversions, and high parallelism. This research focuses on building an all-optical DCT processor based on the DNN architecture (DCT–DNN). Testing revealed that this processor performed DCT operations on random matrices and achieved DCT-based compression on specific data sets. Additionally, the DCT with a block for large-sized images was validated. The DCT–DNN, with its high speed and low energy consumption, can be integrated with other complex optoelectronic computing systems to serve as a general computing device for computational acceleration. Furthermore, it can be combined with data transmission systems or directly integrated into image information collection systems to encode and transmit front-end collected information. This makes it a valuable tool for data processing, encryption, and transmission applications.