Metasurface-based All-optical Image Processing

The demand for image information processing in artificial intelligence has surged.At present,digital image processing has the disadvantages of slow speed and high energy consumption,which cannot meet the requirement for the higher intelligent systems.Because massless photons are bosons with low propagation loss at the speed of light,all-optical image processing is one of the potential ways to improve decisionmade speed in artificial intelligence.However,the available all-optical image processing systems exist problems such as single functionality,low integration,and low image quality,which need to be further studied systematically.To solve these problems in this thesis,the structure,complex amplitude and coherence modulation characteristics of geometric phase metasurfaces is used to focus on multi-functional multiplexing,integrated image convolution,image homogenization in all-optical image processing systems.The achievements in this thesis are image encryption,single-chip all-optical image convolution,and speckle-free in holographic image reconstruction.The achieved results in this thesis are simplified as follows:1.To realize multi-dimensional encryption of optical images,double(amplitude and phase)anisotropies of artificial dielectric nanostructures are investigated by using the directionally excited magnetic resonances for developing geometric dielectric(Si,Si3N4)metasurfaces.Based on the independent modulation of optical amplitude and phase,three optical images are encrypted in a single metasurface device by multiplexing orthogonal channels such as the polarization state,coherence state and the wavelength of photons.It provides technological support for multifunctional multiplexing of all-optical images.2.For the problems of single functionality and large volume in current all-optical image processing,complex-amplitude metasurface devices are developed by 3×3 convolution kernel for modifying the point spread function of an optical imaging system,so as to realize arbitrary all-optical image convolution operations in such an optical imaging system.Experimentally a single-chip or double-chip all-optical image convolution system has been demonstrated with high-quality edge detection,spatial differentiation,denoising and edge enhancement.The optical convolution kernel has a 3×3 matrix format,which is highly compatible with electronic convolution operations and therefore pays a solid way for developing all-optical convolutional neural networks.3.To remove the speckles in holographic image,a neural network algorithm is used to optimize the modulation function of optical spatial coherence and the phase function of holographic imaging simultaneously,yielding a phase hologram for speckle-free image theoretically.Experimentally,the expected phase profile is realized by using the geometric metasurface,which creates an holographic image with high uniformity and sharper edges(enhanced by 2 orders of magnitude compared with the previous results)under the illumination of high-coherence laser.It proves the superiority of the holographic design based on neural network algorithm.Furthermore,it is used to demonstrate the holographic lithography that creates the forked grating and two-dimensional code with good performance,thus providing a new insight for speckle-free holographic image reconstruction.