deep-wood: automated wood species identification using ... tuo he.pdf · more than 10 specimens per...

37
1.Wood Anatomy and Utilization Department Research Institute of Wood Industry Chinese Academy of Forestry He Tuo 1,2 , Prabu Ravindran 3,4 , Lu Yang 1,2 , Alex C. Wiedenhoeft 3,4 , Jiao Lichao 1,2 , Yin Yafang 1,2* 3. Center for Wood anatomy Research Forest Products Laboratory United States Department of Agriculture 2. Wood Collections (WOODPEDIA) Chinese Academy of Forestry IAWA-IUFRO Symposium Beijing 2019 Deep-wood: Automated wood species identification using convolutional neural networks

Upload: others

Post on 15-Aug-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

1.Wood Anatomy and Utilization Department

Research Institute of Wood Industry

Chinese Academy of Forestry

He Tuo1,2, Prabu Ravindran3,4, Lu Yang1,2, Alex C. Wiedenhoeft3,4, Jiao Lichao1,2, Yin Yafang1,2*

3. Center for Wood anatomy Research

Forest Products Laboratory

United States Department of Agriculture

2. Wood Collections

(WOODPEDIA)

Chinese Academy of Forestry

IAWA-IUFRO Symposium Beijing 2019

Deep-wood: Automated wood species identification using convolutional neural networks

Page 2: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

01 02 03 04 05 06

Outline

Page 3: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

01 02 03 04 05 06

Page 4: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×
Page 5: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

International efforts to combat the illegal logging

Chinese Regulations

on Wild Plants

Protection Protect, develop and

utilize wild plant species

that are listed in the

international treaty and

national regulations

USA Lacey Act

Amendment Prohibits all the trade of

the plants and plant

products that are illegally

sourced from US state and

foreign counties

European Union

Timber Regulation

Prohibits the illegally

harvested timber and

timber products on the

European market

Australia

Illegal Logging

Prohibition Act Prohibits wood, pulp and

paper products into Aus.

or process Aus. raw logs

illegally logged

Page 6: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

7 new timber proposals:

Cedrela,

Pterocarpus tinctorius,

Dalbergia,

Guibourtia,

Pericopsis elata

CITES (Convention on International Trade in Endangered Species of Wild Fauna and Flora, since 1975)

CITES Cop

Total Number of Timber Species in CITES Appendix

CITES Appendix for Timber Species

Ⅰ Ⅱ Ⅲ

2010 Cop15

111 7 94 10

2013, CoP16

247 7

231 New added:Dalbergia cochinchinensis, D. granadillo, Osyris lanceolata and 48 Dalbergia spp. and 84 Diospyros spp.( populations of Madagascar) From III to II:Dalbergia retusa, D. stevensonii

9

2016, Cop17

~500 7

~486 New added: Dalbergia spp., Guibourtia tessmannii, Guibourtia demeusei, Guibourtia pellegriniana From III to II:Pterocarpus erinaceus

8

Page 7: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Standard lists

Reference samples

Wood anatomists

Technical tools

Professional wood

anatomists working with

highly trained ground staff

Access to xylarium

collections and associated

tools for wood anatomical

analysis

Genus level

Page 8: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Image

Computer Vision

Feature

Classifier

Macroscopic

Microscopic

GLCM

Wavelet Transform

Local banalization

BP-neural network

Support Vector Machine

K-nearest neighbor

Page 9: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Input Features

Input Output

Artificially feature engineering

Automated feature representation

Output

Page 10: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

01 02 03 04 05 06

Page 11: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Wood specimens from 4 xylaria ---FPL, CAF, INF and IPT

15 Dalbergia species—218 specimens

Species Quantity of Wood Specimens

MADw/SJRw CAFw SPSFw BCTw Total

Dalbergia cearensis 5 0 2 7 14 Dalbergia cochinchinensis 5 1 1 0 7 Dalbergia frutescens var.tomentosa 8 0 1 7 16 Dalbergia hainanensis 1 1 0 0 2 Dalbergia hupeana 2 7 0 0 9 Dalbergia latifolia 15 2 1 3 21 Dalbergia melanoxylon 9 0 0 2 11 Dalbergia nigra 21 1 8 26 56 Dalbergia odorifera 0 4 0 0 4 Dalbergia oliveri 4 2 0 0 6 Dalbergia retusa 16 0 0 0 16 Dalbergia sissoo 18 0 0 3 21 Dalbergia spruceana 3 0 3 6 12 Dalbergia stevensonii 11 0 0 0 11 Dalbergia tucurensis 12 0 0 0 12

Page 12: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

11 Pterocarpus species—161 specimens

Species Quantity of Wood Specimens

MADw/SJRw CAFw SPSFw BCTw Total

Pterocarpus dalbergioides 12 0 0 0 12

Pterocarpus erinaceus 4 5 0 0 9

Pterocarpus indicus 25 2 1 1 29

Pterocarpus macrocarpus 13 3 1 1 18

Pterocarpus marsupium 13 0 0 2 15

Pterocarpus officinalis 20 0 0 1 21

Pterocarpus rohrii 9 0 0 1 10

Pterocarpus santalinus 4 0 0 0 4

Pterocarpus soyauxii 6 4 5 3 18

Pterocarpus tinctorius 5 1 0 0 6

Page 13: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Sample Polish Image Collection Dataset Creation

Dataset with 10,237 images

Page 14: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Total: 132,265

Train: 105,986

Val: 13,134

Test: 13,145

2048x2048

1600x1600 227x227

Patch dataset creation

Page 15: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

CNN architecture-VGG16 & AlexNet

Page 16: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Transfer learning

www.image-net.org/ Google Deep Mind Training CNNs pre-trained on ImageNet

VGG16

ImageNet dataset

Wood images

Trained CNNs

14,197,122 images, 21,841 synsets

Page 17: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Training & Testing

0

2

4

6

8

10

12

0

15

50

31

00

46

50

62

00

77

50

93

00

10

85

0

12

40

0

13

95

0

15

50

0

17

05

0

18

60

0

20

15

0

21

70

0

23

25

0

24

80

0

26

35

0

27

90

0

29

45

0

31

00

0

32

55

0

34

10

0

35

65

0

Loss

Loss

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

30

0

18

00

33

00

48

00

63

00

78

00

93

00

10

80

0

12

30

0

13

80

0

15

30

0

16

80

0

18

30

0

19

80

0

21

30

0

22

80

0

24

30

0

25

80

0

27

30

0

28

80

0

30

30

0

31

80

0

33

30

0

34

80

0

36

30

0

Accuracy

Accuracy

Stochastic Gradient Descent

Page 18: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

01 02 03 04 05 06

Page 19: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

VGG16-15 Dalbergia species model

Average accuracy: 85.44%

Correct classified species(100%): D. frutescens D. oliveri D. hupeana D. sissoo D. melanoxylon D. nigra D. stevensonii

Totally misclassified species: D. cochinchinensis

Poorly misclassified species: D. odorifera

Page 20: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Leguminosae_Dalbergia_cochinchinensis_CAFw_20373_15567504-2018-05-30-174456.png

Leguminosae_Dalbergia_odorifera_CAFw_19152_15567504-2018-05-22-014340.png

Misclassified images

Standard images

Page 21: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Average accuracy : 67.44%

Correct classified species(100%): P. soyauxii

Poorly misclassified species: P. indicus P. rohrii

VGG16-11 Pterocarpus species model

Page 22: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

P. angolensis P. indicus P. dalbergioides P. macrocarpus P. marsupium P. erinaceus

P. tinctorius P. santalinus P. soyauxii

P. officinalis P. rohrii

Page 23: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

(1.000)

(1.000)

(1.000)

(1.000) (0.900)

(1.000) (0.960)

(0.967) (1.000)

(0.900) (0.767)

(1.000) (0.000)

Page 24: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Relative wood anatomical variability within class

Re

lative w

oo

d an

atom

ical distin

ctness

Low

H

igh P. soyauxii (1.000)

P. officinalis (0.967) P. rohrii (0.067)

P. macrocarpus (0.867) P. marsupium (0.743) P. dalbergioides (0.700) P. erinaceus (0.684)

P. angolensis (0.900) P. indicus (0.367)

P. santalinus (0.577) P. tinctorius (0.657)

Page 25: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

AlexNet:15-species model

Average accuracy: 93.68%

Correct classified species(100%):

Poorly misclassified species: Dalbergia nigra (78.12%)

Dalbergia melanoxylon Dalbergia odorifera

Page 26: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

AlexNet: 11-species model

Average accuracy: 88.38%

Poorly misclassified species: Pterocarpus indicus

Page 27: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

AlexNet: 26-species model

Average accuracy:99.34%

Correct classified species: 12 Dalbergia species 7 Pterocarpus species

Poorly misclassified species: Pterocarpus indicus

Page 28: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

AlexNet

Page 29: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

No. of the specimens per species ≧ 10

Model accuracy ≧ 85%

No. of the images per species ≥ 100 ≥ 300

Model accuracy ≥ 99% Model robustness

Page 30: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Model accuracy: high-quality datasets > low-quality dataset Patch size ≥ 1000×1000

Page 31: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

01 02 03 04 05 06

Page 32: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

AlexNet (93.68%/88.68%) outperforms VGG16 (85.44%/67.44%) on the image dataset

(Dalbergia/Pterocarpus) collected in this study.

Parameters for AlexNet model:

More than 10 specimens per species

Over 100 high-quality images per species

Patch size of 1000 × 1000 × 3

Automated computer vision models for field screening of wood species to combat

illegal logging.

Page 33: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

01 02 03 04 05 06

Page 34: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Global network to establish wood image database

Page 35: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Unpack the "black box" for feature representation in terms of wood anatomy

Page 36: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

-- National Forestry & Grassland Administration (NFGA), China

-- China CITES Management Authority

-- National Natural Science Foundation of China

-- China Scholarship Council

Page 37: Deep-wood: Automated wood species identification using ... Tuo HE.pdf · More than 10 specimens per species Over 100 high-quality images per species Patch size of 1000 × 1000 ×

Wood Anatomy and Utilization Department

Research Institute of Wood Industry

Chinese Academy of Forestry

Center for Wood anatomy Research

Forest Products Laboratory

United States Department of Agriculture

Wood Collections

(WOODPEDIA)

Chinese Academy of Forestry

IAWA-IUFRO Symposium Beijing 2019

Thanks for Your Comments!