Abstract:Driven by the development of reprogrammable metasurface, we realize the dynamic feature extraction using reprogrammable coding metasurface in this work for the first time. The targets here are images of three-dimensional human body with the constant contrast of 13 (see Fig. 1a), and we could recognize and classify them using the extracted feature. The scenario could be described like that, the target is illuminated by the radiation field of programmable coding metasurface, and then the scattered field of the target would be recorded by a single detector. The fundamental mathematical expression is $\boldsymbol{y}\mathbf{W}^{T}\boldsymbol{x}$ , where the two column vectors $\boldsymbol{x}\sim \Re^{\mathbf{M}\times 1}$ and $\boldsymbol{x}\sim \Re^{\mathbf{N}\times 1}$ represent the target images and received data, respectively, and $\mathbf{W}\sim \Re^{\mathbf{M}\times \mathbf{N}}$ is a dimension-reduction matrix. In fact, each column of $\mathbf{W}$ is a radiation pattern of the metasurface at the target image's position and we need to switch the coding sequence of the programmable metasurface electrically, which is easy and fast, and record the data for $\mathbf{N}$ times. As we known, the entries in $\mathbf{W}$ will affect the quality of classification, and different coding sequence of the metasurface results in different $\mathbf{W}$ , so proper coding sequences of the metasurface need to be carefully manipulated. The reprogrammable metasurface can be trained by using the principal component analysis (PCA), a widely used method in the area of machine learning, when a collection of samples is available (see Fig. 1). As comparison (Fig. 1 and Fig. 2), we compare the classification results obtained by the reprogrammable metasurface trained by the PCA technique and random projection in Figure 2, where the idea results obtained by performing the PCA analysis are also presented. we also try the conventional way, classification after imaging reconstruction to find that it requires more data to guarantee the accuracy of the classification though using some compressive imaging methods. Actually, the average classification accuracy in our work can achieve about 80%, in which some could exceed 90%, and we can improve some numbers' accuracy according to practical needs with a compromise of other numbers. In addition, some influence factors on the classification accuracy like scale of the metasurface, the amount of received data, just the switching times of the metasurface, namely the dimensional-reduction ratio of object images, have been taken into consideration.

High‐Resolution Programmable Metasurface Imager Based on Multilayer Perceptron Network

Machine-learning Reprogrammable Metasurface Imager.

An Enhanced Diffractive Neural Network for Metasurface Holograms with High Resolution, Low Noise, and Uniform Intensity

Single-frequency and Single-sensor Three-dimensional Microwave Imaging Using Programmable Metasurface

High-performance full-color imaging system based on end-to-end joint optimization of computer-generated holography and metalens.

Imaging Resolution Analysis of Single-Frequency and Single-Sensor Programmable Microwave Imager

High Resolution 3D Microwave Real Time Imaging System Based on Intelligent Metasurface

Transmission-Type 2-Bit Programmable Metasurface for Single-Sensor and Single-Frequency Microwave Imaging

Dynamic Feature Extraction Using Reprogrammable Coding Metasurface

Single-shot and single-sensor high/super-resolution microwave imaging based on metasurface

Metasurface-based Spectral Convolutional Neural Network for Matter Meta-imaging

Intelligent Imaging and Sensing based on Programmable Metasurface

Large-aperture computational single-sensor microwave imager using 1-bit programmable coding metasurface at single frequency

Microwave Imaging Enhanced by Duplexed Random Phase Modulations

High‐Resolution Near‐Field Imaging and Far‐Field Sensing Using a Transmissive Programmable Metasurface

On the Application of Deep Learning in Modeling Metasurface

Deep Learning-Assisted Simultaneous Targets Sensing and Super-Resolution Imaging

PISC-Net: A Comprehensive Neural Network Framework for Predicting Metasurface Infrared Emission Spectra

Integrated Photonic Metasystem for Image Classifications at Telecommunication Wavelength.

Noninvasive inset-integrated meta-atom for achieving single-layer metasurface simultaneously with coded microwave reflectivity and digitalized infrared emissivity

Metasurface-enabled on-chip multiplexed diffractive neural networks in the visible