Spectral equalization diffractive neural network for image classification

By introducing rectangular gratings and specific optical path layouts into the diffraction neural network, the input image spectrum is equalized and arrayed, solving the problem of insufficient feature extraction caused by spectrum imbalance, improving classification accuracy, and maintaining the advantages of optical computing. It is suitable for image classification tasks.

CN122242605APending Publication Date: 2026-06-19UNIV OF ELECTRONICS SCI & TECH OF CHINA
0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
UNIV OF ELECTRONICS SCI & TECH OF CHINA
Filing Date
2026-03-09
Publication Date
2026-06-19

AI Technical Summary

⚠Technical Problem

Existing diffraction neural networks based on Fraunhofer diffraction suffer from insufficient feature extraction and limited classification accuracy in image classification tasks due to the uneven distribution of spectral intensity in the input image, making it difficult to meet the requirements of high-precision recognition.

⚗Method used

A rectangular grating is used to achieve spectral equalization of the input image. Combined with a specific optical path layout, the spectral distribution of the input image is controlled through the synergistic effect of the rectangular grating, Fourier lens, pinhole screen and multi-layer diffraction layer, so as to equalize and array it, thereby improving the efficiency of feature extraction.

🎯Benefits of technology

It significantly improved classification accuracy from 95% to 97.2%, while maintaining a compact structure that is easy to apply in industrial applications, and features high bandwidth, low power consumption, and parallel processing capabilities.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

This invention discloses a spectral equalization diffraction neural network for image classification, belonging to the field of optical computing and neural network technology, and applicable to image classification tasks. Addressing the problem of uneven spectral intensity distribution and insufficient high-frequency feature extraction in existing Fraunhofer diffraction neural networks, which limit classification accuracy, this invention employs an optical path structure of "input layer - rectangular grating - Fourier lens - pinhole screen diffraction layer group - output layer." A rectangular grating with a period of 32 pixels achieves spectral equalization and arraying. Combined with three diffraction layers (256*256 diffraction neurons per layer, 4mm interlayer spacing) for step-by-step feature extraction, and a 20mm focal length Fourier lens with a 532nm operating wavelength, after 20 rounds of backpropagation training on the MNIST dataset, the classification accuracy reaches 97.2%, an improvement of nearly 2 percentage points compared to existing technologies. This invention has a compact structure and strong feasibility, inheriting the advantages of all-optical computing—high bandwidth, low power consumption, and parallel processing—enabling efficient all-optical image recognition.
Need to check novelty before this filing date? Find Prior Art