黇 싟髰ꦬ꼤|ꥻꊗ댷驸시쐏 묗驇시꾣ꂃ쐏
8VZJJ_J3JY 鮳ꦛ듛 띃끠 뜠ꛛ 鮳ꨘ듇 쏟됬싟 몋싷깄쏗 鮳ꨘ
꺟뒣ꜿ 馘꺠뙿
꺟馘ꂃ시髓 떇딓驸시驿
ZYOIGKQF%XTLFSLFHPWXOPFSL%XTLFSLFHPW
8VZJJ_J3JYGFXJI8NSLQJ.RFLJ8ZUJW7JXTQZYNTS
ZXNSL0ST\QJILJ)NXYNQQFYNTS
>ZQNR8JT 8ZP/Z0FSL
8TLFSL:SN[JWXNY^
됗 대
魿ꔛ듛 몋싷깄쏗 XZUJWWJXTQZYNTS87 덳髯ꀗ 麧쀻둏뷯 鯍驣 麖騏 ꝏꉧ댷 꺴ꀨ듇 鼕듷ꀗꃳ 뙿 듷ꚛꁧ 魻ꕯ鸛 ꅜ끟덓 鼕듃 덳긳ꕌ驿 ꟗ꠫ꜯ 꼏ꯇꕌ듷 뜠馃싛ꀗ ꢻ떟馃 ꦟ깠싛鮳 ꋏꢻ덓 듷 끧떟ꗟ 싛ꉟ둫댷ꗟ 髯쎇싛鮳덓ꀗ 댷ꖧ됷 ꢻ떟馃 뗷딯싟 ꁧ 魻ꖊ鮳덓 됳ꜯꀗ 麧쀻둏뷯 뫟떄쏗 뼸싷 꺴ꀨ 馓꼏 뫟꼏쏗싛ꟷ꺟 삏꓿ꥻ뺳 꾛 뚇듷ꀗ 麧쀻둏뷯 8VZJJ_J87듇 꺧驇싛 뎃ꁧ ꎓ싟 띃끠 뜠ꛛ 0ST\QJILJ)NXYNQQFYNTS0) 듷됬싷 묗馃떄듻 삏꓿ꥻ뺳 꾛 뜠馃 덉듷 꺴ꀨ듇 鼕듿 꾛 딋ꀗ 시뀸 ꦬꨘ 듇 떟닋싟ꁧ ꎓ싟 0) 끟 YJFHMJWSJY\TWP듛 꺴ꀨ듷 ꩷ꁧ XYZIJSYSJY\TWP덓 딛 떇ꁯꆛꅇꗠ KJFYZWJRFU 馇듛 ꯇ髓 뼸싷 시뀸 쐫뒫듇 鼕듿 꾛 딋덋ꁧ 騳驿떄뒿ꗟ 됳ꜯꀗ 0) 鮳ꨘ듇 뼸싷 묗馃떄듻 삏꓿ꥻ뺳 꾛 뜠馃 덉듷 꺴ꀨ듇 鼕덯 ꁧ 87麧쀻둏 뷯 ꂃꯇ ꃗ ꯣ꛷驣 꺴ꀨ 馓꼏 뫟꼏쏗싟 麧쀻둏뷯 떟닋싟ꁧ꺟ꗣ
몋싷깄쏗 8ZUJW7JXTQZYNTS 87듃 떃싷깄ꅇ 뎄깄뒿ꗟꬃ뺳 驣싷깄ꅇ 뎄깄 ꩸둓싛ꀗ 뱷쉫뺳 ꯇ떇듛 싟 닿듷ꁧ 黇 87(33@B듃 뫟몋ꗟ ꊨꕯꁠ듇 긯됬싛덯 87 듇 꾛쌌싛뎃ꁧ 듷덓 뎄쌨듇 ꦞ닇 ;)87@B 驿 馜듃 ꩷ꁧ 꺴ꀨ듷 鼕듃 麧쀻둏뷯ꉧ듷 듷쑇 꼏馟ꆛ덋ꁧ 魻ꕯ鸛 듷ꕯ싟 麧쀻둏뷯ꉧ듃 ꖋ듷댷꾛馃 ꝑ驣 ꩸딤싛鮳 ꋏꢻ덓 鼕듃 덳긳ꕌ ꦒ ꝑ듃 삏꓿ꥻ뺳 됗髯싛ꀗ ꢻ떟馃 뗷딯싟ꁧ 뒇듛 ꢻ떟 싷騳싛鮳 뒇싷 됳꺣 됳ꜯꀗ 8VZJJ_J3JY@B덓꺟 뎄馓듇 댾댷 꺴ꀨ馓꼏 뫟꼏쏗싛驣 鸱듃 삏꓿ꥻ뺳 꾛 馃띃ꀗ XYZIJSY SJY\TWP 듻 8VZJJ_J87 꺧驇싛뎃驣 띃끠 뜠ꛛ 0ST\QJILJ )NXYNQQFYNTS 0) @B듇 뼸싷 시뀸ꦬꨘ듇 ꩃ驀싛덯 ꩷ꁧ 鼕듃 꺴ꀨ듇 馃띃ꅇꗠ 麧쀻둏뷯 시뀸싷꩷닛ꁧ 듷ꋏ 꺴ꀨ듇 鼕듷鮳 뒇싷 ꅜ끟 시뀸덓긯됬싟 YJFHMJW SJY\TWP ꗟꀗ 87 麧쀻둏뷯 뚔 8YFYJTKYMJ FWY 8TYF 麧쀻둏뷯 뚔 싛鸛듻 *)87 듇 鮳ꦛ뒿ꗟ 꺧驇싟 *)87 9 긯됬싛뎃ꁧ
떟닋싛ꀗ ꦬꨘ
麧쀻둏뷯 髯뗳 魻ꜿ 덓꺟 쏘듻싣 꾛 딋ꉲ듷 YJFHMJW SJY\TWP 듻 *)879 ꀗ 됳꺣 KJFYZWJ J]YWFHYNTS 듇 뒇싷 (TS[ QF^JW 싛鸛 髯꺴싛뎃驣 듷쑇 7JXNIZFQ 'QTHP 馟 덳騳싛뎃ꁧ 馄馄듛 7JXNIZFQ 'QTHP 듃 몠 馟듛 7JXNIZFQ 2TIZQJ ꗟ 髯꺴ꆛ댷 딋ꁧ 馄馄듛 7JXNIZFQ 2TIZQJ 듃 (TS[ QF^JW 馟 7J1: 馟 魻ꜯ驣 XHFQNSL 듇 뒇싟 2ZQYNUQJ QF^JW 싟 馟ꗟ 髯꺴 ꆛ댷딋ꁧ 馄馄듛 7JXNIZFQ 2TIZQJ 듃 ꠫ꇓ XPNU HTSSJHYNTS 뒿ꗟ 덳騳ꆛ댷 딋驣 7JXNIZFQ 'QTHP ꎓ싟 XPNU HTSSJHYNTS 뒿ꗟ 덳騳ꆛ댷 딋ꁧ 8YZIJSY SJY\TWP 듻 8VZJJ_J87 듛 髯뗳ꀗ YJFHMJW SJY\TWP 돃 馜듷 KJFYZWJJ]YWFHYNTS 듇 뒇싷 먛듏덓 (TS[QF^JW 싛鸛 髯꺴 싛뎃驣 듷쑇 8VZJJ_J 'QTHP 馟 덳騳싛뎃ꁧ 8VZJJ_J 'QTHP黇 싟髰ꦬ꼤|ꥻꊗ댷驸시쐏 묗驇시꾣ꂃ쐏 싛鸛ꀗ 몠 ] KNQYJW 馟돃 ] KNQYJW 싛鸛ꗟ 髯꺴ꆛ댷 딋ꁧ 馄 馄듛 8VZJJ_J 'QTHP 듃 ꠫ꇓ XPNU HTSSJHYNTS 뒿ꗟ 덳騳ꆛ댷 딋 ꁧ 鮳뗷듛 8VZJJ_J3JY 驿 ꁧ 떓듃 IT\SXFRUQNSL QF^JW 髯꺴싛띃 닍듃 떓듷ꁧ 띃끠 뜠ꛛ 鮳ꨘ 9JFHMJW SJY\TWP ꗟꬃ뺳 ꩷ꁧ 鸛듃 떘꩷ 댾鮳 뒇싷꺟ꀗ KJFYZWJRFU ꁫ뒇덓꺟 떘꩷ 댾ꀗ 騆듷 싇됗싛ꁧ 魻ꜿ 덓꺟 ꩷듷ꉲ듷 YJFHMJW SJY\TWP 듛 KJFYZWJ RFU 9 9 9 ꀗ 馄馄 7JXNIZFQ 'QTHP 듇 뼸驿싟 묟ꖨ KJFYZWJ RFU 듷驣 XYZIJSY SJY\TWP 듛 KJFYZWJ RFU 8 8 8 ꀗ 8VZJJ_J 'QTHP 듇 뼸驿 싟 묟ꖨ KJFYZWJ RFU 듷ꁧ 馄馄듛 KJFYZWJ RFU 듃 꾟꺟덓 ꝡ 騏 띠뒿ꗟ 髯꺴ꆛ댷 꼓끧 싫꾛듛 딈ꖨ뒿ꗟ 긯됬ꆟꁧ
끧썛 騳驿
듷ꨋ 딨덓꺟ꀗ 떘꺴떄 떘ꕌ떄 끧썛떄 騳驿돃 꺠듇 ꩷덯뚇 騆 듷ꁧ 쇟 덓꺟ꀗ 0) 듇 떄됬싛띃 닍듃 8VZJJ_J87 驿 0) 떄 됬싟 8VZJJ_J870) 魻ꜯ驣 YJFHMJW SJY\TWP 듻 *)879 돃 떘ꕌ떄 騳驿 ꯇ髓 쏘듻싣 꾛 딋ꁧ 8VZJJ_J87 듃 ꁧ 87 ꦬ ꨘ ꂃꯇ ꝧ됳 鸱듃 삏꓿ꥻ뺳 꾛 싇됗ꗟ 싛띃ꝏ 꺴ꀨ 馓꼏 셰듷 뷯띃 닍듃 騆듇 쏘듻싣 꾛 딋ꁧ ꎓ싟 *)879 듛 띃ꅇ 닇ꔛ 시 뀸끟븫 8VZJJ_J870) 듛 驀됳 묗馃떄듻 삏꓿ꥻ뺳 꾛 뜠馃 덉듷 꺴ꀨ듷 뎯꓿馇 騆듇 쏘듻싣 꾛 딋ꁧ ꎓ싟 듷ꀗ 'NHZGNH ꩷馇ꨘ 驿 뒣긯싟 꾛뚃듛 꾛쌌끟馇듇 馃띇ꁧ 魻ꜿ 덓꺟ꀗ ꁧ 87 鮳 ꨘꉧ驿 떘꺴떄 ꯇ髓 쏘듻싣 꾛 딋ꁧ 묟ꖨ 듷ꥻ띃 ꩷ꟷ 쏘듻 싣 꾛 딋ꉲ듷 떟닋ꆟ ꦬꨘ듃 ꩷ꁧ 꺣ꠈ싟 뎄깄듇 묟ꖨ싛ꀗ 騆듇 닏 꾛 딋ꁧ
騳ꗣ
됳ꜯꀗ 麧쀻둏뷯 뫟떄쏗 鮳ꨘ듇 뼸싷 ꁧ 87 麧쀻둏뷯덓 ꯇ싷 鸱듃 삏꓿ꥻ뺳꾛 馃띃驣 덳긳 꼐ꅇ馃 ꯣ꛷ꟷ꺟 꺴ꀨ馓꼏 뫟 꼏쏗싟 8VZJJ_J87 듇 꺧驇싛뎃驣 0) 鮳ꨘ듇 쏟됬싷 싷ꁼ 麧쀻 둏뷯듛 꺴ꀨ듇 鼕듷ꀗ 시뀸 鮳ꨘ듇 떟끟싛뎃ꁧ ꎓ싟 떘ꕌ떄 떘 꺴떄 騳驿 ꠫ꇓ덓꺟 쏘듻싣 꾛 딋ꉲ듷 꺴ꀨ 馓꼏셰덓 ꯇ싷 싷ꁼ 麧쀻둏뷯馃 ꁧ 麧쀻둏뷯덓 ꯇ싷 ꝧ됳 ꝑ듷 驀ꕌ쏗ꆟ 騆듇 쏘듻 싣 꾛 딋ꁧ릻뗳
[1] Dong, Chao, et al. "Image super-resolution using deep convolutional networks." IEEE transactions on pattern analysis and machine intelligence 38.2 (2015): 295-307.
[2] J. Kim, J.K. Lee, and K.M. Lee. “Accurate image super-resolution using very deep convolutional networks.” IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1637-1645, 2016.
[3] F. N. Iandola, M. W. Moskewicz, K. Ashraf, S. Han, W. J. Dally, and K. Keutzer. “Squeezenet: Alexnet-level accuracy with 50x fewer parameters and¡ 1mb model size.” arXiv preprint arXiv:1602.07360, 2016.
[4] G. Hinton, O. Vinyals, and J. Dean. “Distilling the knowledge in a neural network.” arXiv preprint arXiv:1503.02531, 2015.
긯긯
ꩻ 덳髯ꀗ 黇ꅇ 떘ꬃ 驿시鮳꾣떘꩷뼸끣ꬃ듛 딯둓뒿ꗟ 싟髰 덳髯딯ꁫ듛 띃둓 3T 2-&& 驿시鮳꾣떘꩷뼸 끣ꬃ ꦒ 떘꩷뼸끣鮳쐐셌馃둓듛 ꂃ시 .(9 덳髯꺿뺳띃둓긯덈듛 덳髯 騳驿 ..95 ꦒ ꢻ쏗먷뒤骃骔ꬃ ꦒ 싟髰볛뻓 뮣띇쓨둓듛 덳髯馟ꦟ띃둓긯덈뒿ꗟ 꾛쌌ꆛ덋듏 驿떟ꨋ쎻 7 쇟 셌馃 ꃳ듷뺳껎듇 쏟됬싟 떘ꕌ떄 騳驿Dataset Scale Bicubic EDSR
-T Squeeze SR Squeeze SR-KD Set5 2 33.66 37.89 36.62 36.90 3 30.39 34.17 32.33 32.70 4 28.42 31.75 30.67 30.86 Set14 2 30.24 33.28 32.34 32.56 3 27.55 29.97 28.97 29.24 4 26.00 28.20 27.63 27.70 Urban 100 2 3 26.88 24.46 31.29 27.43 29.43 25.71 29.60 26.10 4 23.14 25.33 24.59 24.73
(a) Ground truth
(PSNR/time) (b) Bicubic (22.28/-) (29.70/2.7734) (c) SRCNN
(c) EDSR-T
(31.68/2.4623) (d) SqueezeSR (27.62/0.2667) (e) SqueezeSR-KD (29.62/0.2384)
̐ս 2. Urban100 ࢂ “img_004” ࠒۘࢂ ࢽ۽ࢶ ʼ˕ ̐ս1. Knowledge Distillation ࡶ ࡢଞ ˱࣏
