Zithunzi zobalalika zopanda pamtima zotengera ma ultrafast convolutional neural network

Spatial light modulator ndi mtundu wa chigawo champhamvu chomwe chimatha kusinthira matalikidwe, gawo ndi polarization mkhalidwe wa kuwala kochitika munthawi yeniyeni moyang'aniridwa ndi chizindikiro chakunja. Kugwiritsa ntchito modulator yowunikira malo m'munda wa Kujambula Kwamwaza sangagwiritsidwe ntchito popanga kuwala kwa pseudo-thermal kuwala m'malo mwa galasi lachikhalidwe, komanso angagwiritsidwe ntchito ngati cholinga chomwaza kafukufuku wazithunzi. Kugwiritsa ntchito ma modulator amlengalenga kumatha kuzindikira zoyambira ndikuwongolera pakuwongolera malo amwazikana.

Zambiri za Thesis:

Zotsatira za kukumbukira kwa kuwala kwakhala maziko ofunikira a njira zoyerekeza zazikulu komanso zazing'ono muzofalitsa zovuta zobalalika, kuphatikiza minyewa yamtambo ndi zigawo zamathothomathotho. Komabe, kukonzanso zithunzi muzofalitsa zofalitsa kwambiri popanda kuwala kwa kukumbukira sikunapambane. Kuti izi zitheke, tikuwonetsa kuthekera kopanganso zithunzi m'malo obalalika popanda zotsatira za kukumbukira popanga multistage convolutional optical neural network (ONN) yomwe imaphatikiza ma cores angapo omwe amagwira ntchito pa liwiro la kuwala. Kutengera ndi Fourier Optics, parallel, sitepe imodzi convolution maphunziro a ONN, mawonekedwe amachotsedwa mwachindunji kuti athe kukonzanso zithunzi popanda kukumbukira, ndi Field Of View imakulitsidwa mpaka nthawi 271. Chipangizochi chitha kusinthidwanso kuti chimangidwenso mwachangu kwambiri pamitundu yambiri ndi mphamvu yapakompyuta ya 1.57peta pa sekondi imodzi (POPS), ndikukhazikitsa bwino makina ophunzirira makina othamanga kwambiri komanso ogwira mtima opangira zithunzi.

Zotsatirazi ndi gawo la zoyeserera ndi zotsatira:
Kukonzekera koyesera kwa convolutional ONN kukuwonetsedwa mu Chithunzi 1. Poyesera, chowonetsera mphamvu modulator SLM0 (FSLM-HD70-A/P, kukula kwa pixel 8μm, 1920 × 1080) amagwiritsidwa ntchito popanga kulowetsa kwa pixel-level ya convolutional ONN. Munda wowunikira womwe umasinthidwa ndi SLM0 umatumizidwa kudzera mu dongosolo la 4f kupita ku ndege yolowera ya convolution ONN. Dongosolo la 4f lili ndi magalasi awiri okhala ndi Kutalika kwa Focals wa 100 mm (L1) ndi 50 mm (L2) kuti apereke 0.5x kukula kwa zinthu zolembedwa ngati SLM0. Ndege yolowetsayo imatumizidwa ku ONN yokhala ndi magawo atatu kuti ikonzedwenso. Convolution ONN imakhala ndi ma module atatu a SLM (chitsanzo chathu FSLM-4K70-P02). Gawo SLMs ali 4094 × 2400 mapikiselo ndi mapikiselo kukula ndi 3.74μm×3.74μm. Gawo loyamba la convolution limasungidwa ngati SLM1, ndipo mtunda pakati pa ndege yolowera ndi SLM1 ndi 10cm. Pambuyo posinthidwa ndi SLM1, mtandawo umatumizidwa ku SLM2 ndi kusinkhasinkha kuchokera ku ma splitter a BS2 ndi BS3. Chigawo chachiwiri cha convolution chimasungidwa ngati SLM2, ndipo SLM1 ndi SLM2 ndi 20cm. Mtengo wosinthidwa ndi SLM2 umatumizidwa ku SLM3. Chigawo chachitatu cholumikizidwa kwathunthu chimayikidwa mu SLM3, ndi mtunda wa 10 cm pakati pa SLM2 ndi SLM3. Mtsinje wopangidwa ndi SLM3 umafalitsidwa kupita ku ndege ya convolution ya ONN mwa kuwonetsera kuchokera pazitsulo za BS4, ndi mtunda wa 20cm pakati pa SLM3 ndi ndege yotuluka. Kamera yogulitsira (kukula kwa pixel 4.8μm, 1280 × 1024) idayikidwa pa ndege ya convolutional ONN kuti ijambule zotsatira zake. Ndikofunika kuzindikira kuti kusagwirizana pakati pa zigawo zosiyana za convolution ONN kungawononge kwambiri ntchito yake.

CHITH. Chithunzi cha 1 choyeserera choyeserera cha magawo atatu a convolutional network ONN. (A) Phatikizani magawo ndi zolumikizira za gawo lililonse la ONN. (B) Zida zoyesera zosinthira ONN. M, galasi; POL, polarizer yozungulira; BS, mtengo woboola pakati; L, mandala; SLM, malo opangira kuwala.

CHITH. 2 Makina omangiriranso zithunzi zopanda kukumbukira kudzera mumwaza wosanjikiza. (A) Chithunzi chobalalika kudzera m'mabala angapo. Wobalalitsa aliyense akhoza kutsatiridwa ngati wosanjikiza wochepa thupi, ndi N kuyimira chiwerengero cha omwaza ndi d kuyimira kusiyana pakati pa omwaza. (B) Mapangidwe a timadontho-thonga opangidwa pansi pamikhalidwe yosiyana yobalalika. Chithunzi chakumanzere chikufanizira mawanga omwe ali ndi mawonekedwe okumbukira (N = 1, d = 0) ndi mawanga opanda zotsatira za kukumbukira (N> 1, d> 0). Chithunzi chomwe chili kumanja chikuwonetsa zokhotakhota zokhotakhota pazochitika zilizonse, ndi mizere yeniyeni yoyimira ma curve aang'ono amtundu umodzi wagalasi. Theka la kutalika ndi m'lifupi mwa ma curve awa zimatsimikizira malo owonera chithunzicho. (C) Lingaliro lakumanganso zithunzi zopanda kukumbukira mu ONN ya magawo atatu.

CHITH. 3 Mfundo ya Optical convolutional neural network. (A) ONN yosanjikiza itatu imakhala ndi zigawo ziwiri zolumikizirana ndi mawonekedwe amodzi olumikizidwa kwathunthu. (B) Mapangidwe a gawo loyamba la convolution ali ndi mitundu isanu ndi inayi yosiyana. Chigawo chilichonse chimakhala ndi zigawo zitatu: gawo la vortex, gawo losasintha, ndi gawo la grating. (C) Gawo la gawo lachiwiri la convolution lagawidwa m'zigawo za 3 × 3. Gawo la dera lililonse limapangidwa ndi njira yofanana ndi gawo loyamba la convolution, zomwe zimapangitsa kuti pakhale ma nuclei 81. (D) Kuchita kwachindunji kwa ONN kophunzitsidwa m'magulu azinthu zotengera MNIST ndi ma dataset a FashionMNIST. Chithunzi pamwambapa chikuwonetsa mayendedwe ophunzirira pazochitika zilizonse. Chithunzi chotsatirachi chikuwonetsa zotsatira za kuyesa kwa ndege yotulutsa ya ONN, yokhala ndi mabwalo ofiira odumphira omwe akuyimira dera lozindikira lophunzitsidwa ndi nambala yofananira.

CHITH. 4 Kutsimikizira koyeserera pakumanganso zithunzi zopanda kukumbukira. (A) Zotsatira zoyeserera zoyeserera za convolution yophunzitsidwa ya ONN yokhala ndi zigawo ziwiri zobalalika (N = 2). Mzere woyamba umasonyeza chithunzi chenicheni cha chinthucho, ndipo mzere wachiwiri mpaka wachinayi umagwirizana ndi zotsatira zomanganso d = 3 cm, d = 4 cm, ndi d = 5 cm, motero. (B) Zotsatira zoyeserera zoyeserera za ma ONN ophunzitsidwa bwino okhala ndi zigawo zingapo zobalalika (N> 2). Mzere woyamba mpaka wachinayi umagwirizana ndi zomwe N = 2, 3, 4, ndi 5, motsatana. (C) Zotsatira za gawo la vortex pa convolution ONN zikuwonetsedwa. Chithunzi chakumanja chikufanizira ma curve kernel okhala ndi gawo la vortex pomwaza d = 5 cm ndi zotsatira zomanganso. Chithunzi chakumanzere chikuwonetsa kuti zida zosiyanasiyana za gawo la vortex zitha kugwiritsidwa ntchito kutulutsa zidziwitso zakumbali za chinthucho.

Chithunzi 5 chikuwonetsa machitidwe amphamvu komanso ochita zambiri a ma ONN ophunzitsidwa bwino. (A) Dynamic inference process (S1 ndi S2) pomanganso zithunzi zopanda kukumbukira pogwiritsa ntchito convolutional ONN. Njira yolowera imayikidwa pa 60Hz SLM. (B) Dongosolo lachiwonetsero la ONN la ntchito ziwiri likuwonetsedwa mu S3 pamawonekedwe azithunzi zamakanema ambiri kuti athe kukonzanso zithunzi popanda kukumbukira. Chiwonetsero cha gawo lachitatu la raster yolumikizidwa kwathunthu chikuwonetsedwa Chithunzi S17.

Magawo amtundu wa matalikidwe owunikira kuwala kogwiritsidwa ntchito poyesaku ndi awa:

Magawo a gawo-mtundu wa spatial light modulator omwe amagwiritsidwa ntchito poyesera ndi awa:

Lembani kumapeto:
Pankhani ya kuyerekeza kwa computational, ma neural neural network agwiritsidwa ntchito kwambiri kuthetsa mavuto pazithunzithunzi za ghost, digito holography, Fourier lamination microscopy ndi zina. Nthawi yomweyo, kuthekera kokwanira kwa data komanso kukhathamiritsa kwa njira yophunzirira mwakuya kumathandizanso kwambiri pantchito yofalitsa zithunzi. Ndi kusintha kwabwino komanso kuwongolera kolondola kwa modulator yowunikira kuwala, kuphatikiza kwa spatial light modulator ndi optical neural network kutulutsa zowala zambiri.

Zambiri zankhani:
DOI: 10.1126/sciadv.adn2205