In optical computing systems for image tasks,optical signals of image data are usually generated by signal modulators,whose speed limits computational performance.To address this,a framework of image memory that can p...In optical computing systems for image tasks,optical signals of image data are usually generated by signal modulators,whose speed limits computational performance.To address this,a framework of image memory that can permanently store the image and instantaneously generate its corresponding optical signals has been proposed.In this framework,an original image is sliced into non-overlapping sub-images,each of which contains 3 pixel×3 pixel or 5 pixel×5 pixel,and these subimages are respectively stored in storage units,which are nine-channel or 25-channel power splitters(PSs).The gray values of pixels in each sub-image are sequentially mapped to the light intensity of channels in each PS.These PSs are inversely designed by combining the adjoint method(AM)and the finite-difference frequency-domain method.Using this method,high-performance memory has been realized based on PSs,with a maximum 33.37 dB signal-to-noise ratio.Using multiobjective AM,the storage capacity has been doubled according to the working wavelengths.For example,two sub-images with pixels arranged as‘T'and‘L'shapes can be retrieved at 1310 and 1550 nm,respectively.This framework provides highly stable,non-volatile image storage,and instantaneous generation of optical signals for future all-optical computing systems.展开更多
基金supported by the National Natural Science Foundation of China(No.62161025)the High-level Talent Introduction Research Program of Shanghai University of Engineering Science(No.2023RC-GC09)the High-Level Local University Construction Innovative Talents Training Program of Shanghai University of Engineering Science(No.25XAI008)。
文摘In optical computing systems for image tasks,optical signals of image data are usually generated by signal modulators,whose speed limits computational performance.To address this,a framework of image memory that can permanently store the image and instantaneously generate its corresponding optical signals has been proposed.In this framework,an original image is sliced into non-overlapping sub-images,each of which contains 3 pixel×3 pixel or 5 pixel×5 pixel,and these subimages are respectively stored in storage units,which are nine-channel or 25-channel power splitters(PSs).The gray values of pixels in each sub-image are sequentially mapped to the light intensity of channels in each PS.These PSs are inversely designed by combining the adjoint method(AM)and the finite-difference frequency-domain method.Using this method,high-performance memory has been realized based on PSs,with a maximum 33.37 dB signal-to-noise ratio.Using multiobjective AM,the storage capacity has been doubled according to the working wavelengths.For example,two sub-images with pixels arranged as‘T'and‘L'shapes can be retrieved at 1310 and 1550 nm,respectively.This framework provides highly stable,non-volatile image storage,and instantaneous generation of optical signals for future all-optical computing systems.
