Pattern recognition system based on a coherent diffractive correlator with deep learned processing of downsampled correlation responses
Deep convolutional neural networks are known for high precision of object recognition; however, processing of high-resolution images with the use of high-resolution kernels requires a lot of calculations during training and inference. Optical Fourier-processors and correlators provide highly paralle...
Gespeichert in:
Veröffentlicht in: | Applied optics (2004) 2024-12, Vol.63 (36), p.9196 |
---|---|
Hauptverfasser: | , , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | Deep convolutional neural networks are known for high precision of object recognition; however, processing of high-resolution images with the use of high-resolution kernels requires a lot of calculations during training and inference. Optical Fourier-processors and correlators provide highly parallel calculations that are robust to electromagnetic interference and potentially energy efficient. Article results demonstrate that the correlation pattern recognition problem can be efficiently solved by implementation of deep neural network for processing of downsampled output signals of coherent diffractive correlators. The results of neural network-based correlation processor architecture study, numerical training, and experimental implementation are presented and discussed in the article. It is shown that output signals of optical correlators being captured by a low-resolution sensor can be efficiently classified by a deep neural network that was trained on a numerically generated laboratory database of correlation responses. The use of auto-correlation peak-narrowing techniques such as phase modulation and contouring of input images or application of optimized distortion-invariant filters allow us to unify the form of auto-correlation peaks such that there is no need for retraining of the network if the target object is changed. Application of three trained network models with input layer sizes of 32×32, 16×16, and 8×8 for processing the downsampled correlation responses of different experimental implementations of 4-f and 1-f coherent diffractive correlators optoelectronic schemes, which include the schemes based on binary spatial light modulation, proved the possibility to perform recognition of objects on 256×256 images with precision above 92% and potential processing speed of more than 1000 frames per second. |
---|---|
ISSN: | 1559-128X 2155-3165 |
DOI: | 10.1364/AO.541305 |