Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Attention-Based Adaptive Selection of Operations for Image Restoration in the Presence of Unknown Combined Distortions [CVPR 2019 読み会]

773 views

Published on

複数のdistortionを含む画像に対しても適用可能なCNNアーキテクチャの提案.

Published in: Technology
  • Be the first to comment

Attention-Based Adaptive Selection of Operations for Image Restoration in the Presence of Unknown Combined Distortions [CVPR 2019 読み会]

  1. 1. CC6 C 2B6 2 C D6 6 6 C 6A2C B A 2 6 .6BC A2C C96 -A6B6 6 0 3 6 BC AC B I 1-.
  2. 2. • • ) [ V • A O LIR +1 ]LIDB E LC • ( A O B E LC • = + . G M 8 7 )9 P . G M 81 2 (9 P . 80 , 9 @
  3. 3. - - - - - - - - - - - - - • , 3 3 • A Input Ours Ground truth Mix Raindrop Blur Noise JPEG SE R T UA PXC I MKLNE O
  4. 4. - • • 4 - - - -
  5. 5. D • B • 5 • +
  6. 6. ,,K QHN • . 6 • VRP ILM 2 0 2 0 + C 2 C 0 , 219 8
  7. 7. C C NO P , ILE A K D I EI DI E I ED E C IE I ED • D DE 0 D C D 0 D 1 A M A A I D • 1E EA 2D K IM , • 9 E A 7 • 5
  8. 8. 2 8 810 1 , 1 2 8 810 881 82 1 , 1
  9. 9. 9: I ( op1 op2 op3 op4 op1 op2 op3 op4 … Input op1 op2 op3 op4 )1 2
  10. 10. D A F • 3 03 • 3 • 0 3 1 3 1 -
  11. 11. • 2 ( ) 6 • 2 ( ) 12 3×3
  12. 12. • 3 - )2 3×3 • ( 1 ) (
  13. 13. - FeatureExtraction Block Operation-wise AttentionLayer Operation-wise AttentionLayer Reconstruction Layer Operation-wise AttentionLayer • • 1 1 3 3
  14. 14. AC + 4 18C 7E 8 8 F 8 CC 01 . ,E 2AE88H8 7 8 F C 01
  15. 15. ! ∈ ℝ$ • 5%&'( ∈ ℝ)×+×$ - ! ∈ ℝ$ • 1
  16. 16. !" = [%&" ' , … , %&" * ] • ,- ∈ ℝ0×2, ,3 ∈ ℝ|*|×05" 6" = ,78(,'6") %&" ; = exp(5"(6")) ∑ exp(5"(6")) = exp(&" ; ) ∑;@' |*| exp(&" ; ) → B C 6 1
  17. 17. ! "# • $%&'(')*% +,-. 1×1 • $%&'(')*% +,-. 3×3 • $%&'(')*% +,-. 5×5 • $%&'(')*% +,-. 7×7 • $%&'(')*% 45*'6%4 +,-. 3×3 (( = 2) • $%&'(')*% 45*'6%4 +,-. 5×5 (( = 2) • $%&'(')*% 45*'6%4 +,-. 7×7 (( = 2) • '.%(';% &,,*5-; 3×3
  18. 18. !" #" $×$ . 1×1 ℱ(: 1×1 *+,-
  19. 19. • - = 1 • 4 9 = • = Feature maps !×#×$ %&'( )& *+,-( Feature maps !×#×$ %& Feature maps !×#×$ %&-,'( … OperationLayerAttentionLayer … Concat 1×1conv. /& OperationLayer Concat 1×1conv. *+,-,
  20. 20. • - -- - 2 0 - Feature maps !×#×$ %&'( )& *+,-( Feature maps !×#×$ %& Feature maps !×#×$ %&-,'( … OperationLayerAttentionLayer … Concat 1×1conv. /& OperationLayer Concat 1×1conv. *+,-,
  21. 21. • 2 • 1 3 1 23 1 3 , • 2 + • 1 3 1 23 1 3 , • 2 • 1 1 1 3 1 23 1 3 , 3 312
  22. 22. , • D : 8 : : • 63×63 C I: PYV R[V K V ] • : + C 8 : : : + C 8 : • 2 b • a 4 5 9 ] , , : - :9 4 5 3 , 8 : : 9::C : 8: : : . ,201
  23. 23. 02.1 22 D IK 16 8 76 3 , 54 4 6 EP PE 6I D 6MDE E :ESE E 6E IC 8:79 ::36 8:79 ::36 8:79 ::36 1L077 ) ) ) ) ) )- ( 95 9EP M E , ( () )) ) )- ) 4 L E ETMLD RPPI L DELMIPE . 9EPIDR E LIL M DEEN CLL M IK E DELMIPIL 3222 LP C IMLP ML 3K E 8 MCEPPIL . ( )) 4 R E 0 IL MM C IL M IK E EP M IML BT DEEN EIL M CEKEL E LIL 3L 0 89 , • v tnr 077edf • v uo 8 E ILED 077cighm]aYZp VZs bl [ u DIP M IML l
  24. 24. 1 4 - 4 2- 2 42 4
  25. 25. PS IG 5 CD 5D R G PSV JL O • 1(3)( 0) 5 • 33 A D 5 • ,5ECC 5 C ,5ECC 5 E 1 , 7 A CC 5 D 57DC ( CD D C 2 CD 0E D 7D
  26. 26. ] V [cedf kgn 73 7 LMIK ob a • h mi P I K LMIK M I Yl aYR 182785 6 1 76 .41 01 1 376 /6 KI K I M IMM L K (. ) ) ) , . ( ( ( . )- ,) . ,- % % , . % ) ), - % (, (. - ( , ( . , ) ( ) ,( ) ,( - ) ) , , . ( % % % M C K IP M I CIKL KLI M LC LIA MK MO (( , . ) ( )( . ) )) - ( () ) - - ( ) , ,- - - ( % . , , - % - , % ) - % ) - ( (. ( ( ) , )) , () , )( ) ( (( - ). - . - ) - - - . . , , ( , . ( ) ( ( , . % % % 850 , 89 73 7 LMIK P I K LMIK M I 89 73 7 LMIK P I K LMIK M I 4 MCI 2 9 M 0K AM MII C AIK K LMIK M I K AIK M K 1 0867 -
  27. 27. Cat Cat Cat Sheep Person Person Person Person Boat Dog Dog Person Cow Cow Dog Dog Dog Dog Person Person Person Person Person Dog Bottle Bottle Clean Distorted RL-Restore [1] Ours Cat Person Person Person Person Person Person Person Person Cat TVmonitor Person Person Person Person Horse Horse Horse HorseCar
  28. 28. I JE A G • .2 2 2A5 0 1 IRQ • 128×128 5 I • 2 A5 2 385A 5A A5 2 385A • ( , ) I P ) 0 1 . 2 5 2 5 C5 5 5 2 C5 2 C5 A2 2 5 D : 2 5 C2 2 A 5 2 5 ,. 52 A 5
  29. 29. • S - . 2 / • - . 2 >9 - - 9 D • - - - - • a - G A
  30. 30. • RE N g Ga e i J • 1 h RE N • 33 3 33 3 P d J 1 E B P G
  31. 31. : 9 58 2 199A () (- . -. 4429 TA) () . .( 3 103A ( - .(+ 9 ( 4 351 04 89 48 : 9 58 2 199A ( ) -- , 8 N9 TA+ ( - -,- 9 . 4. 84 1 9168 .5 .9 42. 9168 .5 SK IM FKSTPRTKP h] bifkald [c eg A) 6 C I 4 F T P CRF C CST C F M KDM SPM TKP PR E DCS F KNCI F PKSK I 5333 RC SCETKP S P 5NCI :RPE SSK I , . )+ -Z)+ - A 8 IC NC 3 MPKTK I T PT TKCM P STC FCRF EP VPM TKP CM C TP EPF RS PR KNCI R STPRCTKP DY VPM TKP CRY S CRE 5 5187 - A+ CK 8 N T 0 RSKST T N NPRY T PRL PR KNCI R STPRCTKP 5 511 ,
  32. 32. • 2CC C a • t gec gecVy ] R • C V s • kbfapKy 0 1 Y I- V r R • u 2CC C a R - 2 3 Y]K ldm n hio c P [ VN L[ 0 1 C 2 2 C 8 2 C 2 28 C 2C 3 C 2 8 .,-
  33. 33. • • - - • e r p • D V I • - 43 l I n • - d 4 - ? e o m • • b g h d 4 a - - i • - -

×