A Custom CNN Approach for Detection and Classification of Image artefacts in Silicon Hyperbolic Metamaterials

Hyperbolic metamaterials (HMMs) possess a hyperbolic dispersion profile in Fourier space (k-space), enabling deep sub-wavelength imaging over much of the electromagnetic spectrum. However, existing theoretical analysis and experimental demonstrations have shown that imaging with HMMs suffers from unwanted imaging artefacts that restrict their operation for broad-band imaging. In this paper, we study the image transmission properties of silicon-based HMMs at THz frequencies. From our numerical simulation, we show that the proposed metamaterial can resolve feature sizes below the diffraction limit but suffers from severe image artefacts. Using a convolutional neural network (CNN) we classify between good and bad images and also use a custom SqueezeNet model architecture to detect the good images. The proposed techniques can classify the good images with an accuracy of about 98%.