際際滷

際際滷Share a Scribd company logo
Unet
レ郁規 螳豌 覿 覈(UNet)  螳
Image Segmentation
Detection / Segmentation
Panoptic: 覈 蟆   覲伎企, 碁朱
螳豌 蟆豢 (Object Detection)  螳豌伎 豺, 譬襯 蠏
襴螻 襯 豢ロ.
[x, y, width, height, class, probability]
企語 覿 (Image Segmentation)  蟆  螳
 3螳讌襦 蟲覿.
 Semantic Segmentation (覩碁 覿)
 Instance Segmentation (語ろ伎 覿, 螳豌 覿)
 Panoptic Segmentation
Semantic Segmentation
Semantic Segmentation(覩碁 覿) 螳 Class  螳 豸″ 覦覯. ( る れ)
 螳 螳豌(Same class, different instance)螳 蟆轟 朱 蟲覿  .
螳覦 讌螳 覦 讌企化 覯
覈
Instance Segmentation
Instance Segmentation(語ろ伎 覿, 螳豌 覿) 讌 螳豌企ゼ Instance 襦 蟆豢. ( る れ)
 螳 螳豌(Same class, different instance)螳 蟆轟 企 蟲覿  .
Panoptic Segmentation
Panoptic Segmentation Semantic Segmentation + Instance Segmentation 企.
 螳 螳豌願 蟆轟 企 蟲覿  螻 覦郁化  蟲覿  .
Segmentation 譬襯
覲 企語 Panoptic Segmentation
Semantic Segmentation
Instance Segmentation
Object Detection
Classification (覿襯)
Binary Classification
(伎 覿襯)
Saliency Detection
(譴 覓殊牡 蟆豢)
Semantic Segmentation
(覩碁 覿)
Multi-Label Classification
(れ 覿襯)
Object Detection
(螳豌 蟆豢)
Instance Segmentation
(語ろ伎 覿, 螳豌 覿)
Panoptic Segmentation
ex) ResNet, EfficientNet
ex) FasterRCNN, YOLO
ex) MaskRCNN, YOLACT
ex) U^2 Net,
Pyramid Feature Attention
ex) U-Net, HRNet, DeepLab
ex) UPSNet, Mask2Former
U-Net 蟲譟 (豢 伎)
Output Mask 蠍磯   蠍一 狩. 豈 Class 螳企.
Input Image(3 x 256 x 256) -> Output Mask( [c] x 256 x 256 )
U-Net 覈
Q)  Encoder 覿覿  蠍瑚?
Encoder Decoder
殊 蠏碁殊 るジ讓 襷ろ襦 蟆所襯
谿場 .
Binary Cross Entropy Loss
log(0) (覓危)企襦  る襯 覦讌蠍     (epsilon)朱 覯襯 螻  螻.
巨巨巨巨巨 = 
1

鐃
=1

 鐃 log  + 1   鐃 log 1   , ゐゐゐゐゐゐゐゐ   ,   1  
Ground Truth ()
Prediction (豸)
糾骸 豸′伎伎 曙 螳朱 る 螻.
糾骸 豸′ るゴ覃 (覓危) 螳蟾讌螻, 螳朱 0 螳蟾 讌り 危危覃 暑.
 螻殊
   .
曙朱Μ 觜蟲 蟆襷 蠍一牛!
Binary Cross Entropy Loss 螳 蠍一 る 螻 覦覯.
Binary Cross Entropy Loss
巨巨巨巨巨 = 
1

鐃
=1

 鐃 log  + 1   鐃 log 1   , ゐゐゐゐゐゐゐゐ   ,   1  
蟆讀 覺
y螳 1  p螳 0 螳蟾讌覃(讀, 襷 襴 襦) 覓危 螳蟾讌.
y螳 0  p螳 1 螳蟾讌覃(讀, 襷 襴 襦) 覓危 螳蟾讌.
畉芯‘i勅[畉ナ弼i勅gi寧[畉弼[
畉芯‘i勅[棐p[畉弼[
畉梶坤[a畉6棐pp3[04
[[[[p[[2ag畉親p3[2畉ツ釈弼2畉a畉ナ畉2畉テ弼畉gi03[
裡[9[2畉ツ釈弼2畉a畉ナ畉2畉テ弼畉gi00
[[[[[[2ag畉親3[2畉ツ釈弼2畉a畉ナ畉2畉テ弼畉gi03[
裡[9[2畉ツ釈弼2畉a畉ナ畉2畉テ弼畉gi00
[[[[勅畉ヅ椗[9p[[2gi0[/[裡[9[p0[[2gi裡[9[002帖畉メ0
畉芯[癸-畉帖畉モ-[癸財[高-帖畉畉進癸-4
[[[[p[[2畉勅勅畉p3痢3[裡3[痢3[裡40
[[[[[[2畉勅勅畉p3痢3[痢2立3[痢2痢痢裡3[痢240
[[[[6勅畉ドg[[2畉ツ釈弼2gi弼弼畉ド2畉進畉勅p納勅i弼弼畉ナ勅ip
勅畉メ些a畉i2畉ツ釈弼2gi弼弼畉ド2畛畉メ些a畉i2畍按潤鍵0p3[02棐p0
[[[[6勅畉ドg[[a畉6棐pp3[0
[[[[print[a畉4[16勅畉ドg20
[[[[print[a畉4[16勅畉ドg20
[a畉4[痢2裡陸律里痢陸率率離里立率率立里離率
[a畉4[痢2裡陸律里痢律陸痢律離律立率率離
勅ia畉ド弼[畉進畉辛辿畉メ些[寧畉薪辿[畉ヅ畉薪[ai畉梶坤[痢
Inference (豢襦)
log(0) (覓危)企襦  る襯 覦讌蠍     (epsilon)朱 覯襯 螻  螻.
糾骸 豸′伎伎 曙 螳朱 る 螻.
糾骸 豸′ るゴ覃 (覓危) 螳蟾讌螻, 螳朱 0 螳蟾 讌り 危危覃 暑.
豢襦 螻殊
讌 threshold(覓誤煙) 伎 1襦 覩碁 0朱 覲蟆!
 Encoder 覿覿 企 Backbone  覓企逢.
  企語蠍一 1/32螳   蟾讌 Convolution一 .
( 襦 譟一 螳, TF2 Backbone 1/32蟾讌 譴 危 cls-head襯 覿企  1/32蟾讌 )
 Width, Height螳 譴企り鍵 螳 襷讌襷 伎伎 燕 襯 蟇一 Tensor襯 Concatenate .
 InceptionResNet螳 Convolution一一 strid paddin朱 蟆 覦朱 譴企れ  蟆曙磯 Concatenate  豢螳 padding  .
U-Net 蟲譟 (Encoder)
Encoder (Backbone)
U-Net 蟲譟磯手鍵 覲企 U-Net Encoder襯 襷 覦覯 !
ゐゐゐゐゐゐゐゐ
2
,

2
ゐゐゐゐゐゐゐゐ
4
,

4
ゐゐゐゐゐゐゐゐ
8
,

8
ゐゐゐゐゐゐゐゐ
16
,

16
ゐゐゐゐゐゐゐゐ
32
,

32
ゐゐゐゐゐゐゐゐ
1
,

1
覲 Encoder螳企殊語 Tensorflow2 Backbone 蠍磯朱 朱,   蟲 るゼ  .
ResNet50 ( 256 x 256 ) 蟆曙 Tensor 蠍磯 ( BatchSize x 8 x 8 x 2048 )
Pooling  企 蟇 讌 Chanel 蠍郁 企れ 譴讌 . Backbone螻 螳 讌 れる 讌讌 蟯 !
U-Net 蟲譟 (Decoder)
Decoder (Upsample or Transposed Conv)
覲 Encoder螳企殊語 Tensorflow2 Backbone 蠍磯朱 朱,   蟲 るゼ  .
Output Layer 螳 pixel 0 ~ 1  螳 螳讌.
Threshold襯 牛 Binary Mask襯 襷れ伎 .
ゐゐゐゐゐゐゐゐ
2
,

2
ゐゐゐゐゐゐゐゐ
4
,

4
ゐゐゐゐゐゐゐゐ
8
,

8
ゐゐゐゐゐゐゐゐ
16
,

16
ゐゐゐゐゐゐゐゐ
32
,

32
ゐゐゐゐゐゐゐゐ
1
,

1
 Encoder Concatenate  伎企ゼ (Upsample, ConvTr) れ 覿碁.
  企語 蠍郁   蟾讌 一 .
  蠍磯 襴  螻, 蠍磯ゼ 襴螻 Residual螻 螳 伎企 豢螳襦  螳.
U-Net 蟲譟 (Decoder::Upsample, ConvTr)
Upsample / Transposed Conv
Transposed Convolution 覦襯 覈豺, Deconvolution 覈視 覈豺.
 企語 resize 觜訣.
 豢
 Stride ( レ  伎 蟇磯Μ ) 2, kernel_size  3 蟆曙.
U-Net 蟲譟 (Concatenated Skip Connection)
U-Net Encoder Decoder
 螳 ъ ( element wise add ) 螳  Tensor襯 覿碁 !
ゐゐゐゐゐゐゐゐ
4
,

4
ゐゐゐゐゐゐゐゐ
8
,

8
ゐゐゐゐゐゐゐゐ
16
,

16
ゐゐゐゐゐゐゐゐ
32
,

32
ゐゐゐゐゐゐゐゐ
1
,

1
ゐゐゐゐゐゐゐゐ
2
,

2
ゐゐゐゐゐゐゐゐ
2
,

2
ゐゐゐゐゐゐゐゐ
4
,

4
ゐゐゐゐゐゐゐゐ
8
,

8
ゐゐゐゐゐゐゐゐ
16
,

16
ゐゐゐゐゐゐゐゐ
32
,

32
ゐゐゐゐゐゐゐゐ
1
,

1
 ResNet skip-connection Add讌襷, U-Net Concatenate朱 谿願 .
 り浬 狩 覲企ゼ れ 伎伎    .
Adobe Photoshop (Instance Segmentation)
Adobe Photoshop CC 23.0 螳豌 谿剰鍵 蟲  燕 .
Semantic Segmentation 覈
Badrinarayanan, Vijay, Alex Kendall, and Roberto Cipolla. "Segnet: A deep convolutional encoder-decoder architecture for image segmentation." IEEE transactions on pattern analysis and machine intelligence 39.12 (2017): 2481-2495.
Long, Jonathan, Evan Shelhamer, and Trevor Darrell. "Fully convolutional networks for semantic segmentation." Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.
Sun, Ke, et al. "High-resolution representations for labeling pixels and regions." arXiv preprint arXiv:1904.04514 (2019).
SegNet FCN
HRNet
U^2 Net
Qin, Xuebin, et al. "U2-Net: Going deeper with nested U-structure for salient object detection." Pattern Recognition 106 (2020): 107404.
RSU (ReSidual Ublock)
Dataset
Polygon Mask
Dataset
Polygon Mask
螳企 蟲覃   螳豌企 企蟾?

More Related Content

Similar to [UNET]Segmentation model, a representative UNet, and a slide for understanding (20)

Decision tree
Decision treeDecision tree
Decision tree
Jeonghun Yoon
Anomaly Detection based on Diffusion
Anomaly Detection based on DiffusionAnomaly Detection based on Diffusion
Anomaly Detection based on Diffusion
ssuserbaebf8
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
Lee Seungeun
07. PCA
07. PCA07. PCA
07. PCA
Jeonghun Yoon
october report in korean
october report in koreanoctober report in korean
october report in korean
nao takatoshi
03.12 cnn backpropagation
03.12 cnn backpropagation03.12 cnn backpropagation
03.12 cnn backpropagation
Dea-hwan Ki
Cnn 覦襭
Cnn 覦襭Cnn 覦襭
Cnn 覦襭
譬 豕
Recurrent Neural Net 企螻 る
Recurrent Neural Net 企螻 るRecurrent Neural Net 企螻 る
Recurrent Neural Net 企螻 る
覦 蟾
貊 れ旧朱 危危 瑚概讌
貊 れ旧朱 危危 瑚概讌貊 れ旧朱 危危 瑚概讌
貊 れ旧朱 危危 瑚概讌
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
jdo
伎 豈覺 觜 螳覦 3殊姶
伎  豈覺 觜 螳覦 3殊姶伎  豈覺 觜 螳覦 3殊姶
伎 豈覺 觜 螳覦 3殊姶
Taekyung Han
貂觜襾誤 襾語 Kevin Murphy Machine Learning Statistic
貂觜襾誤 襾語 Kevin Murphy Machine Learning Statistic貂觜襾誤 襾語 Kevin Murphy Machine Learning Statistic
貂觜襾誤 襾語 Kevin Murphy Machine Learning Statistic
讌 譟
Variational AutoEncoder(VAE)
Variational AutoEncoder(VAE)Variational AutoEncoder(VAE)
Variational AutoEncoder(VAE)
螳覩手記 螳覩手記
れ襦 蟆 谿
れ襦 蟆 谿れ襦 蟆 谿
れ襦 蟆 谿
谿
Deep Learning Into Advance - 1. Image, ConvNet
Deep Learning Into Advance - 1. Image, ConvNetDeep Learning Into Advance - 1. Image, ConvNet
Deep Learning Into Advance - 1. Image, ConvNet
Hyojun Kim
Anomaly Detection with GANs
Anomaly Detection with GANsAnomaly Detection with GANs
Anomaly Detection with GANs
覦 蟾
Chapter 10 sequence modeling recurrent and recursive nets
Chapter 10 sequence modeling recurrent and recursive netsChapter 10 sequence modeling recurrent and recursive nets
Chapter 10 sequence modeling recurrent and recursive nets
KyeongUkJang
Lecture 4: Neural Networks I
Lecture 4: Neural Networks ILecture 4: Neural Networks I
Lecture 4: Neural Networks I
Sang Jun Lee
瑚概讌, 蠍郁 蠏碁Μ螻 ル
瑚概讌, 蠍郁 蠏碁Μ螻 ル瑚概讌, 蠍郁 蠏碁Μ螻 ル
瑚概讌, 蠍郁 蠏碁Μ螻 ル
Jinwon Lee
03. linear regression
03. linear regression03. linear regression
03. linear regression
Jeonghun Yoon
Anomaly Detection based on Diffusion
Anomaly Detection based on DiffusionAnomaly Detection based on Diffusion
Anomaly Detection based on Diffusion
ssuserbaebf8
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
Lee Seungeun
october report in korean
october report in koreanoctober report in korean
october report in korean
nao takatoshi
03.12 cnn backpropagation
03.12 cnn backpropagation03.12 cnn backpropagation
03.12 cnn backpropagation
Dea-hwan Ki
Cnn 覦襭
Cnn 覦襭Cnn 覦襭
Cnn 覦襭
譬 豕
Recurrent Neural Net 企螻 る
Recurrent Neural Net 企螻 るRecurrent Neural Net 企螻 る
Recurrent Neural Net 企螻 る
覦 蟾
貊 れ旧朱 危危 瑚概讌
貊 れ旧朱 危危 瑚概讌貊 れ旧朱 危危 瑚概讌
貊 れ旧朱 危危 瑚概讌
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
jdo
伎 豈覺 觜 螳覦 3殊姶
伎  豈覺 觜 螳覦 3殊姶伎  豈覺 觜 螳覦 3殊姶
伎 豈覺 觜 螳覦 3殊姶
Taekyung Han
貂觜襾誤 襾語 Kevin Murphy Machine Learning Statistic
貂觜襾誤 襾語 Kevin Murphy Machine Learning Statistic貂觜襾誤 襾語 Kevin Murphy Machine Learning Statistic
貂觜襾誤 襾語 Kevin Murphy Machine Learning Statistic
讌 譟
れ襦 蟆 谿
れ襦 蟆 谿れ襦 蟆 谿
れ襦 蟆 谿
谿
Deep Learning Into Advance - 1. Image, ConvNet
Deep Learning Into Advance - 1. Image, ConvNetDeep Learning Into Advance - 1. Image, ConvNet
Deep Learning Into Advance - 1. Image, ConvNet
Hyojun Kim
Anomaly Detection with GANs
Anomaly Detection with GANsAnomaly Detection with GANs
Anomaly Detection with GANs
覦 蟾
Chapter 10 sequence modeling recurrent and recursive nets
Chapter 10 sequence modeling recurrent and recursive netsChapter 10 sequence modeling recurrent and recursive nets
Chapter 10 sequence modeling recurrent and recursive nets
KyeongUkJang
Lecture 4: Neural Networks I
Lecture 4: Neural Networks ILecture 4: Neural Networks I
Lecture 4: Neural Networks I
Sang Jun Lee
瑚概讌, 蠍郁 蠏碁Μ螻 ル
瑚概讌, 蠍郁 蠏碁Μ螻 ル瑚概讌, 蠍郁 蠏碁Μ螻 ル
瑚概讌, 蠍郁 蠏碁Μ螻 ル
Jinwon Lee
03. linear regression
03. linear regression03. linear regression
03. linear regression
Jeonghun Yoon

More from Bomm Kim (6)

ML(KNN, Decision Tree), DL(RCNN, YOLO) educational slide
ML(KNN, Decision Tree), DL(RCNN, YOLO) educational slideML(KNN, Decision Tree), DL(RCNN, YOLO) educational slide
ML(KNN, Decision Tree), DL(RCNN, YOLO) educational slide
Bomm Kim
[FreivaldsAlgorithm]The Freivalds Algorithm is an algorithm that determines w...
[FreivaldsAlgorithm]The Freivalds Algorithm is an algorithm that determines w...[FreivaldsAlgorithm]The Freivalds Algorithm is an algorithm that determines w...
[FreivaldsAlgorithm]The Freivalds Algorithm is an algorithm that determines w...
Bomm Kim
[Miller-Rabin] Detailed explanation of the Miller-Rabin prime number test and...
[Miller-Rabin] Detailed explanation of the Miller-Rabin prime number test and...[Miller-Rabin] Detailed explanation of the Miller-Rabin prime number test and...
[Miller-Rabin] Detailed explanation of the Miller-Rabin prime number test and...
Bomm Kim
SegmentTree Datastructure description and implementation slides
SegmentTree Datastructure description and implementation slidesSegmentTree Datastructure description and implementation slides
SegmentTree Datastructure description and implementation slides
Bomm Kim
BOJ10547_STUDENTSKO_LIS(Longest Increasing Subsequence) Algorithm
BOJ10547_STUDENTSKO_LIS(Longest Increasing Subsequence) AlgorithmBOJ10547_STUDENTSKO_LIS(Longest Increasing Subsequence) Algorithm
BOJ10547_STUDENTSKO_LIS(Longest Increasing Subsequence) Algorithm
Bomm Kim
BOJ4743_ConvexHull,Polygon Algorithm and C++ code
BOJ4743_ConvexHull,Polygon Algorithm and C++ codeBOJ4743_ConvexHull,Polygon Algorithm and C++ code
BOJ4743_ConvexHull,Polygon Algorithm and C++ code
Bomm Kim
ML(KNN, Decision Tree), DL(RCNN, YOLO) educational slide
ML(KNN, Decision Tree), DL(RCNN, YOLO) educational slideML(KNN, Decision Tree), DL(RCNN, YOLO) educational slide
ML(KNN, Decision Tree), DL(RCNN, YOLO) educational slide
Bomm Kim
[FreivaldsAlgorithm]The Freivalds Algorithm is an algorithm that determines w...
[FreivaldsAlgorithm]The Freivalds Algorithm is an algorithm that determines w...[FreivaldsAlgorithm]The Freivalds Algorithm is an algorithm that determines w...
[FreivaldsAlgorithm]The Freivalds Algorithm is an algorithm that determines w...
Bomm Kim
[Miller-Rabin] Detailed explanation of the Miller-Rabin prime number test and...
[Miller-Rabin] Detailed explanation of the Miller-Rabin prime number test and...[Miller-Rabin] Detailed explanation of the Miller-Rabin prime number test and...
[Miller-Rabin] Detailed explanation of the Miller-Rabin prime number test and...
Bomm Kim
SegmentTree Datastructure description and implementation slides
SegmentTree Datastructure description and implementation slidesSegmentTree Datastructure description and implementation slides
SegmentTree Datastructure description and implementation slides
Bomm Kim
BOJ10547_STUDENTSKO_LIS(Longest Increasing Subsequence) Algorithm
BOJ10547_STUDENTSKO_LIS(Longest Increasing Subsequence) AlgorithmBOJ10547_STUDENTSKO_LIS(Longest Increasing Subsequence) Algorithm
BOJ10547_STUDENTSKO_LIS(Longest Increasing Subsequence) Algorithm
Bomm Kim
BOJ4743_ConvexHull,Polygon Algorithm and C++ code
BOJ4743_ConvexHull,Polygon Algorithm and C++ codeBOJ4743_ConvexHull,Polygon Algorithm and C++ code
BOJ4743_ConvexHull,Polygon Algorithm and C++ code
Bomm Kim

[UNET]Segmentation model, a representative UNet, and a slide for understanding

  • 2. Image Segmentation Detection / Segmentation Panoptic: 覈 蟆 覲伎企, 碁朱 螳豌 蟆豢 (Object Detection) 螳豌伎 豺, 譬襯 蠏 襴螻 襯 豢ロ. [x, y, width, height, class, probability] 企語 覿 (Image Segmentation) 蟆 螳 3螳讌襦 蟲覿. Semantic Segmentation (覩碁 覿) Instance Segmentation (語ろ伎 覿, 螳豌 覿) Panoptic Segmentation
  • 3. Semantic Segmentation Semantic Segmentation(覩碁 覿) 螳 Class 螳 豸″ 覦覯. ( る れ) 螳 螳豌(Same class, different instance)螳 蟆轟 朱 蟲覿 . 螳覦 讌螳 覦 讌企化 覯 覈
  • 4. Instance Segmentation Instance Segmentation(語ろ伎 覿, 螳豌 覿) 讌 螳豌企ゼ Instance 襦 蟆豢. ( る れ) 螳 螳豌(Same class, different instance)螳 蟆轟 企 蟲覿 .
  • 5. Panoptic Segmentation Panoptic Segmentation Semantic Segmentation + Instance Segmentation 企. 螳 螳豌願 蟆轟 企 蟲覿 螻 覦郁化 蟲覿 .
  • 6. Segmentation 譬襯 覲 企語 Panoptic Segmentation Semantic Segmentation Instance Segmentation
  • 7. Object Detection Classification (覿襯) Binary Classification (伎 覿襯) Saliency Detection (譴 覓殊牡 蟆豢) Semantic Segmentation (覩碁 覿) Multi-Label Classification (れ 覿襯) Object Detection (螳豌 蟆豢) Instance Segmentation (語ろ伎 覿, 螳豌 覿) Panoptic Segmentation ex) ResNet, EfficientNet ex) FasterRCNN, YOLO ex) MaskRCNN, YOLACT ex) U^2 Net, Pyramid Feature Attention ex) U-Net, HRNet, DeepLab ex) UPSNet, Mask2Former
  • 8. U-Net 蟲譟 (豢 伎) Output Mask 蠍磯 蠍一 狩. 豈 Class 螳企. Input Image(3 x 256 x 256) -> Output Mask( [c] x 256 x 256 ) U-Net 覈 Q) Encoder 覿覿 蠍瑚? Encoder Decoder 殊 蠏碁殊 るジ讓 襷ろ襦 蟆所襯 谿場 .
  • 9. Binary Cross Entropy Loss log(0) (覓危)企襦 る襯 覦讌蠍 (epsilon)朱 覯襯 螻 螻. 巨巨巨巨巨 = 1 鐃 =1 鐃 log + 1 鐃 log 1 , ゐゐゐゐゐゐゐゐ , 1 Ground Truth () Prediction (豸) 糾骸 豸′伎伎 曙 螳朱 る 螻. 糾骸 豸′ るゴ覃 (覓危) 螳蟾讌螻, 螳朱 0 螳蟾 讌り 危危覃 暑. 螻殊 . 曙朱Μ 觜蟲 蟆襷 蠍一牛! Binary Cross Entropy Loss 螳 蠍一 る 螻 覦覯.
  • 10. Binary Cross Entropy Loss 巨巨巨巨巨 = 1 鐃 =1 鐃 log + 1 鐃 log 1 , ゐゐゐゐゐゐゐゐ , 1 蟆讀 覺 y螳 1 p螳 0 螳蟾讌覃(讀, 襷 襴 襦) 覓危 螳蟾讌. y螳 0 p螳 1 螳蟾讌覃(讀, 襷 襴 襦) 覓危 螳蟾讌. 畉芯‘i勅[畉ナ弼i勅gi寧[畉弼[ 畉芯‘i勅[棐p[畉弼[ 畉梶坤[a畉6棐pp3[04 [[[[p[[2ag畉親p3[2畉ツ釈弼2畉a畉ナ畉2畉テ弼畉gi03[ 裡[9[2畉ツ釈弼2畉a畉ナ畉2畉テ弼畉gi00 [[[[[[2ag畉親3[2畉ツ釈弼2畉a畉ナ畉2畉テ弼畉gi03[ 裡[9[2畉ツ釈弼2畉a畉ナ畉2畉テ弼畉gi00 [[[[勅畉ヅ椗[9p[[2gi0[/[裡[9[p0[[2gi裡[9[002帖畉メ0 畉芯[癸-畉帖畉モ-[癸財[高-帖畉畉進癸-4 [[[[p[[2畉勅勅畉p3痢3[裡3[痢3[裡40 [[[[[[2畉勅勅畉p3痢3[痢2立3[痢2痢痢裡3[痢240 [[[[6勅畉ドg[[2畉ツ釈弼2gi弼弼畉ド2畉進畉勅p納勅i弼弼畉ナ勅ip 勅畉メ些a畉i2畉ツ釈弼2gi弼弼畉ド2畛畉メ些a畉i2畍按潤鍵0p3[02棐p0 [[[[6勅畉ドg[[a畉6棐pp3[0 [[[[print[a畉4[16勅畉ドg20 [[[[print[a畉4[16勅畉ドg20 [a畉4[痢2裡陸律里痢陸率率離里立率率立里離率 [a畉4[痢2裡陸律里痢律陸痢律離律立率率離 勅ia畉ド弼[畉進畉辛辿畉メ些[寧畉薪辿[畉ヅ畉薪[ai畉梶坤[痢
  • 11. Inference (豢襦) log(0) (覓危)企襦 る襯 覦讌蠍 (epsilon)朱 覯襯 螻 螻. 糾骸 豸′伎伎 曙 螳朱 る 螻. 糾骸 豸′ るゴ覃 (覓危) 螳蟾讌螻, 螳朱 0 螳蟾 讌り 危危覃 暑. 豢襦 螻殊 讌 threshold(覓誤煙) 伎 1襦 覩碁 0朱 覲蟆!
  • 12. Encoder 覿覿 企 Backbone 覓企逢. 企語蠍一 1/32螳 蟾讌 Convolution一 . ( 襦 譟一 螳, TF2 Backbone 1/32蟾讌 譴 危 cls-head襯 覿企 1/32蟾讌 ) Width, Height螳 譴企り鍵 螳 襷讌襷 伎伎 燕 襯 蟇一 Tensor襯 Concatenate . InceptionResNet螳 Convolution一一 strid paddin朱 蟆 覦朱 譴企れ 蟆曙磯 Concatenate 豢螳 padding . U-Net 蟲譟 (Encoder) Encoder (Backbone) U-Net 蟲譟磯手鍵 覲企 U-Net Encoder襯 襷 覦覯 ! ゐゐゐゐゐゐゐゐ 2 , 2 ゐゐゐゐゐゐゐゐ 4 , 4 ゐゐゐゐゐゐゐゐ 8 , 8 ゐゐゐゐゐゐゐゐ 16 , 16 ゐゐゐゐゐゐゐゐ 32 , 32 ゐゐゐゐゐゐゐゐ 1 , 1 覲 Encoder螳企殊語 Tensorflow2 Backbone 蠍磯朱 朱, 蟲 るゼ . ResNet50 ( 256 x 256 ) 蟆曙 Tensor 蠍磯 ( BatchSize x 8 x 8 x 2048 ) Pooling 企 蟇 讌 Chanel 蠍郁 企れ 譴讌 . Backbone螻 螳 讌 れる 讌讌 蟯 !
  • 13. U-Net 蟲譟 (Decoder) Decoder (Upsample or Transposed Conv) 覲 Encoder螳企殊語 Tensorflow2 Backbone 蠍磯朱 朱, 蟲 るゼ . Output Layer 螳 pixel 0 ~ 1 螳 螳讌. Threshold襯 牛 Binary Mask襯 襷れ伎 . ゐゐゐゐゐゐゐゐ 2 , 2 ゐゐゐゐゐゐゐゐ 4 , 4 ゐゐゐゐゐゐゐゐ 8 , 8 ゐゐゐゐゐゐゐゐ 16 , 16 ゐゐゐゐゐゐゐゐ 32 , 32 ゐゐゐゐゐゐゐゐ 1 , 1 Encoder Concatenate 伎企ゼ (Upsample, ConvTr) れ 覿碁. 企語 蠍郁 蟾讌 一 . 蠍磯 襴 螻, 蠍磯ゼ 襴螻 Residual螻 螳 伎企 豢螳襦 螳.
  • 14. U-Net 蟲譟 (Decoder::Upsample, ConvTr) Upsample / Transposed Conv Transposed Convolution 覦襯 覈豺, Deconvolution 覈視 覈豺. 企語 resize 觜訣. 豢 Stride ( レ 伎 蟇磯Μ ) 2, kernel_size 3 蟆曙.
  • 15. U-Net 蟲譟 (Concatenated Skip Connection) U-Net Encoder Decoder 螳 ъ ( element wise add ) 螳 Tensor襯 覿碁 ! ゐゐゐゐゐゐゐゐ 4 , 4 ゐゐゐゐゐゐゐゐ 8 , 8 ゐゐゐゐゐゐゐゐ 16 , 16 ゐゐゐゐゐゐゐゐ 32 , 32 ゐゐゐゐゐゐゐゐ 1 , 1 ゐゐゐゐゐゐゐゐ 2 , 2 ゐゐゐゐゐゐゐゐ 2 , 2 ゐゐゐゐゐゐゐゐ 4 , 4 ゐゐゐゐゐゐゐゐ 8 , 8 ゐゐゐゐゐゐゐゐ 16 , 16 ゐゐゐゐゐゐゐゐ 32 , 32 ゐゐゐゐゐゐゐゐ 1 , 1 ResNet skip-connection Add讌襷, U-Net Concatenate朱 谿願 . り浬 狩 覲企ゼ れ 伎伎 .
  • 16. Adobe Photoshop (Instance Segmentation) Adobe Photoshop CC 23.0 螳豌 谿剰鍵 蟲 燕 .
  • 17. Semantic Segmentation 覈 Badrinarayanan, Vijay, Alex Kendall, and Roberto Cipolla. "Segnet: A deep convolutional encoder-decoder architecture for image segmentation." IEEE transactions on pattern analysis and machine intelligence 39.12 (2017): 2481-2495. Long, Jonathan, Evan Shelhamer, and Trevor Darrell. "Fully convolutional networks for semantic segmentation." Proceedings of the IEEE conference on computer vision and pattern recognition. 2015. Sun, Ke, et al. "High-resolution representations for labeling pixels and regions." arXiv preprint arXiv:1904.04514 (2019). SegNet FCN HRNet
  • 18. U^2 Net Qin, Xuebin, et al. "U2-Net: Going deeper with nested U-structure for salient object detection." Pattern Recognition 106 (2020): 107404. RSU (ReSidual Ublock)