Wood Defect Detection Based on Depth Extreme Learning Machine - MDPI
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
applied
sciences
Article
Wood Defect Detection Based on Depth Extreme
Learning Machine
Yutu Yang 1 , Xiaolin Zhou 2 , Ying Liu 1, *, Zhongkang Hu 3 and Fenglong Ding 1
1 College of Mechanical & Electronic Engineering, Nanjing Forestry University, Nanjing 210037, China;
yangyutu@njfu.edu.cn (Y.Y.); dfl@njfu.edu.cn (F.D.)
2 School of Artificial Intelligence, Hezhou University, Hezhou 542899, China; zhouxiaolinzxl@hotmail.com
3 Nanjing Fujitsu Nanda Software Technology Co., Ltd., Nanjing 210012, China; huzk.fnst@cn.fujitsu.com
* Correspondence: liuying@njfu.edu.cn
Received: 18 September 2020; Accepted: 22 October 2020; Published: 24 October 2020
Abstract: The deep learning feature extraction method and extreme learning machine (ELM)
classification method are combined to establish a depth extreme learning machine model for wood
image defect detection. The convolution neural network (CNN) algorithm alone tends to provide
inaccurate defect locations, incomplete defect contour and boundary information, and inaccurate
recognition of defect types. The nonsubsampled shearlet transform (NSST) is used here to preprocess
the wood images, which reduces the complexity and computation of the image processing. CNN is
then applied to manage the deep algorithm design of the wood images. The simple linear iterative
clustering algorithm is used to improve the initial model; the obtained image features are used as
ELM classification inputs. ELM has faster training speed and stronger generalization ability than
other similar neural networks, but the random selection of input weights and thresholds degrades
the classification accuracy. A genetic algorithm is used here to optimize the initial parameters of the
ELM to stabilize the network classification performance. The depth extreme learning machine can
extract high-level abstract information from the data, does not require iterative adjustment of the
network weights, has high calculation efficiency, and allows CNN to effectively extract the wood
defect contour. The distributed input data feature is automatically expressed in layer form by deep
learning pre-training. The wood defect recognition accuracy reached 96.72% in a test time of only
187 ms.
Keywords: wood defect; CNN; ELM; genetic algorithm; detection
1. Introduction
Today’s wood products are manufactured under increasingly stringent requirements for surface
processing. In developed countries such as Sweden and Finland with developed forest resources,
the comprehensive use rate of wood is as high as 90%. In sharp contrast, the comprehensive use rate of
wood in China is less than 60%, causing a serious waste of resources. With China’s rapid economic
development, people are increasingly pursuing a high-quality life, which will inevitably lead to an
increase in demand for wood and wood products, such as solid wood panels, wood-based panels,
paper and cardboard, and other consumption levels are among the highest in the world. The existing
wood storage capacity and processing level make it difficult to meet the rapid growth demand. The lack
of wood supply and the low use rate have led to the limited development of China’s wood industry.
Therefore, it is necessary to comprehensively inspect the processing quality of logs and boards to
improve the use rate of wood and the quality of wood products.
The nondestructive testing of wood can accurately and quickly make judgments on the physical
properties and growth defects in wood, and nondestructive wood testing and automation can be
Appl. Sci. 2020, 10, 7488; doi:10.3390/app10217488 www.mdpi.com/journal/applsci
Appl. Sci. 2020, 10, 7488 2 of 14
realized. In recent years, the combined application of computer technology along with detection
and control theory has made great progress in the detection of wood defects. In the nondestructive
testing of wood surfaces, commonly used traditional methods include laser testing [1,2], ultrasonic
testing [3–5], acoustic emission technology [6,7], etc. Computer-aided techniques are a common
approach to surface processing, as they are efficient and have a generally high recognition rate [8,9].
Deep learning was first proposed by Hinton in 2006; in 2012, scholars adopted an AlexNet network
based on deep learning to achieve computer vision recognition accuracy of up to 84.7%. Deep learning
prevents dimensionality in layer initialization and represents a revolutionary development in the field
of machine learning [10,11]. More and more scholars are applying deep learning networks in wood
nondestructive testing. He [12] et al. used a linear array CCD camera to obtain wood surface images,
and proposed a hybrid total convolution neural network (Mix-FCN) for the recognition and location of
wood defects; however, the network depth was too deep and required too much calculation. Hu [13]
and Shi [14] used the Mask R-CNN algorithm in wood defect recognition, but they used a combination
of multiple feature extraction methods, which resulted in a very complex model. However, the current
deep learning algorithms still have problems such as inaccurate defect location, incomplete defect
contour and boundary information in the wood defect detection process. To solve the above problems
and effectively meet the needs of wood processing enterprises for wood testing, we carry out the
research of this article.
The innovations of this article are: (1) Simple pre-processing of wood images using nonsubsampled
shear wave transform (NSST), reducing the complexity and computational complexity of image
processing, as the input of convolutional neural network; (2) application of a simple linear iterative
clustering (SLIC) algorithm to enhance and improve the convolutional neural network to obtain a
super pixel image with a more complete boundary contour; (3) use of genetic algorithm to improve
extreme learning machine and classified the obtained image features. Through the above method,
the accuracy of defect detection is improved, and the recognition time is truncated to establish an
innovative machine-vision-based wood testing technique.
2. Materials and Methods
2.1. Wood Defect Original Image Dataset
According to the different processes and causes of solid wood board defects, they are divided into
biohazard defects, growth defects and processing defects. Among them, growth defects and biohazard
defects are natural defects, which have certain shape and structure characteristics, and are also an
important basis for wood grade classification. Generally speaking, solid wood board growth defects
and biohazard defects can be divided into: dead knots, live knots, worm holes, decay, etc. The original
data set used in the experiment in this article is derived from the wood sampling image in the 948
project of the State Forestry Administration (the introduction of the laser profile and color integrated
scanning technology for solid wood panels). When scanning to obtain wood images, the scanning
speed of the scanner is 170 Hz–5000 Hz; Z direction resolution is 0.055 mm–0.200 mm; X direction
resolution is 0.2755 mm–0.550 mm; and color pixel resolution can reach 1 mm × 0.5 mm. The data set
includes 5000 defect maps of pine, fir, and ash. The bit depth of each image is 24, and the specified size
is at the 100*100 pixel level. Part of the defect image is shown in Figure 1.
Appl. Sci. 2020, 10, x FOR PEER REVIEW 3 of 14
Appl. Sci. 2020, 10, 7488 3 of 14
Appl. Sci. 2020, 10, x FOR PEER REVIEW 3 of 14
Figure 1. Common defects of solid wood such as dead-knot, live-knot and decay.
Figure1.1.Common
Figure Commondefects
defects of
of solid
solid wood
wood such
suchas
asdead-knot,
dead-knot,live-knot
live-knotand
anddecay.
decay.
2.2. Optimized Convolution Neural Network
2.2.2.2.Optimized
OptimizedConvolution
ConvolutionNeural NeuralNetwork
Network
This paper proposes an optimized algorithm which uses NSST to preprocess the images
This
followed Thispaper theproposes
bypaper CNN to an
proposes optimized
defect algorithm
an optimized
extract which
algorithm
features from usesimages
which
wood NSSTNSST
uses to
aspreprocess
to preprocess
a preliminary the CNN
images followed
themodel.
images The
by followed
the CNN
simple bytothe
linear CNNdefect
extract
iterative to extract defect
features
clustering (SLIC) features
from woodfrom
super-pixel wood
images images
as as aalgorithm
a preliminary
segmentation preliminary
CNN CNN to
is model.
used model.
The
analyze The the
simple
simple
linear
wood linear clustering
iterative
images iterative clustering
by super-pixel (SLIC) super-pixel
(SLIC)clustering,
super-pixel which segmentation
segmentation
allows algorithm
the algorithm
defects isinused is used
wood to analyze
to analyze
images the the
andwood local
wood
images byimages
super-pixelby super-pixel
clustering, clustering,
which which
allows theallows
defects the in defects
wood
information regarding defects and cracks to be efficiently located. The obtained information is fed in
imageswood and images
local and local
information
information
regarding defectsregarding defects
to beand cracks to be efficiently located.information
The obtained information is initial
fed
back to the initialand cracks
model, which efficiently
enhances located.
the original TheCNN.obtained is fed back to the
back to the initial model,
model, which enhances the original CNN. which enhances the original CNN.
2.2.1. Structure and Characteristics of Convolution Neural Networks
2.2.1.
2.2.1. Structureand
Structure andCharacteristics
Characteristicsof ofConvolution
Convolution NeuralNeural Networks
Networks
The CNN is an artificial neural network algorithm with multi-layer trainable architecture [15].
The TheCNNCNNisisananartificial
artificialneural
neuralnetwork
network algorithm
algorithm with withpoolmulti-layer
multi-layer trainable
trainable architecture
architecture [15].
[15].
It Itgenerally
generally consists
consists of
of an
an input
input layer, excitation
layer, excitation layer,
layer, pool layer,convolution
layer, convolution layer,
layer, and andfullfull
Itconnection
generally consists of an input layer, excitation layer, pool layer, convolution layer, and full connection
connectionlayer. layer.CNNs
CNNshave have many advantagesin
many advantages interms
termsofofimageimage processing
processing applications.
applications. (1) (1) Feature
Feature
layer. CNNs
extraction have many advantages in terms of image processing applications. (1) Feature extraction
extractionand andclassification
classification can can be be combined
combined into intothe thesame
samenetwork
networkstructure
structure andand synchronized
synchronized
and classification
training can be can be combined
achieved, and the into the same
algorithm is network
fully adaptive. structure
(2) Whenand thesynchronized
image size training
is larger, can
training can be achieved, and the algorithm is fully adaptive. (2) When the image size is larger, the the
bedeepachieved, and the algorithm is fully adaptive. (2) When the image size is larger, the deep feature
deepfeature
feature information
information can can be extracted better.
be extracted better. (3) (3)ItsItsunique
uniquenetwork
networkstructurestructure hashas strong
strong
information
adaptability can be extracted better. (3)imageIts unique network structure has strong adaptability to
inthe
adaptabilitytotothe the local
local deformation,
deformation, image rotation,
rotation, image
image translation,
translation, andand other
other changes
changes in the the
local
input deformation, image rotation, image translation, and other changes in the input image. In this
inputimage.
image.In Inthis
thisstudy,
study, each
each pixel in in the
thewood
woodimage imagewas wasconvoluted
convoluted and andthethe defect
defect feature
feature waswas
study, eachby
extracted
extracted pixel in the wood
byexploiting
exploiting theseimage
these CNNwas convolutedThe
characteristics.
characteristics. and
TheCNNthe
CNN defect feature
network
network was extracted
skeleton
skeleton usedused inby
in this thisexploiting
article
article is is
these CNN characteristics.
shownininFigure
shown Figure2.2. The CNN network skeleton used in this article is shown in Figure 2.
Layer
LayerOut
Out
Input
Input Image
Image 3*100*100
3*100*100
Conv7*7,stride=1
Conv7*7,stride=1 64*98*98
,padding=2 64*98*98
,padding=2
RelUs
RelUs
Max pooling
Max pooling 64*49*49
3*3,stride=2 64*49*49
3*3,stride=2
RelUs
RelUs
Conv7*7,stride=1
128*47*47
,padding=2
Conv7*7,stride=1
128*47*47
,padding=2
RelUs
RelUs
Average Pooling
128*24*24
2*2,stride=2
Average Pooling
128*24*24
2*2,stride=2
RelUs
RelUs
FC 1000
FC 1000
Softmax
Classifer
Softmax
Classifer
Figure 2. CNN network skeleton.
Figure2.2.CNN
Figure CNNnetwork
networkskeleton.
skeleton.Appl. Sci. 2020, 10, 7488 4 of 14
2.2.2. Non-Subsampled Shearlet Transform (NSST)
The NSST can represent signals sparsely and optimally, but it also has a strong direction-
sensitivity [16–18]. Therefore, using NSST to preprocess wood images can preserve the defects
feature of wood images. Redundancies in the wood image information are reduced in addition to the
complexity and computation of image processing with the depth learning method.
2.2.3. Simple Linear Iterative Clustering (SLIC)
The CNN uses a matrix form to represent an image to be processed, so the spatial organization
relationship between pixels is not considered—this affects the image segmentation and obscures the
boundary of the defective region of the wood image. The SLIC algorithm can generate relatively
compact super-pixel image blocks after processing a gray or color image. The generated super-pixel
image is compact between pixels, and the edge contour of the image is clear. To this effect, the SLIC
extracts a relatively accurate contour to supplement the feature contour. The SLIC also works with
relatively few initial parameters. Only the number of hyper pixels needed to segment the image must
be set. The algorithm is simple in principle, and has a small calculation range and rapid running
speed. By 2015, the parallel execution speed had reached 250 FPS; it is now the fastest super-pixel
segmentation method available [19].
2.2.4. Feature Extraction
The optimized CNN model proposed in this paper was designed for wood surface feature
extraction. Knots were used as example wood defects (Figure 3a) to test feature extraction via the
following operations. The input image Figure 3a was directly processed by CNN algorithm to
obtain image Figure 3b, which presented local irregularity and nonsmooth edges in the contour after
enlargement [20]. The SLIC algorithm was used to process the input image (Figure 3a) followed by
longitudinal convolution (Figure 3d). The image shown in Figure 3h was obtained after edge removal
and fusion processing. The defect contour features of Figure 3h are substantially clearer compared to
Figure 3b because the segmentation of wood images using CNN, which is expressed in pixels as a
matrix without considering the spatial organization relationship between pixels, affects the end image
segmentation results. The SLIC algorithm instead extracts the wood defect boundary and contour
information from the original image and feeds back the information to the initial segmentation results
of the CNN model.
The above process reduces the redundancy of local image information in addition to the complexity
and computation of the image processing. The pixel-level CNN model method does not accurately
reveal the boundary of the defective region of wood image, but instead indicates only its general
position. SLIC can extract a relatively accurate contour to supplement it and optimize the initial
CNN model. To this effect, the proposed SLIC algorithm-based method improves the defect feature
extraction of wood images over CNN alone.
The input image (Figure 3a) was processed by the NSST algorithm to obtain the image shown in
Figure 3e, then Figure 3f was obtained using the SLIC algorithm. Vertical convolution was carried out to
obtain the image shown in Figure 3g. Figure 3i was obtained after edge removal and fusion processing.
Consider Figure 3i compared to Figure 3h: although the wood defect contour feature extraction effects
are not obvious, using NSST to preprocess the image reduces environmental interference and training
depth to markedly decrease the computation and complexity of the image processing.Appl. Sci. 2020, 10, 7488 5 of 14
Appl.Sci.
Appl. Sci.2020,
2020,10,
10,xxFOR
FORPEER
PEERREVIEW
REVIEW 55of
of14
14
CNN
CNN
(b)
(b)
SLIC
SLIC CNN
CNN
(a)
(a)
NN
SS
SS (c)
(c) (d) (h)
(h)
(d)
TT
SLIC
SLIC CNN
CNN
(e)
(e) (f) (g)
(g) (i)(i)
(f)
Figure3.3.Contrast
Figure Contrastdiagram
Contrast diagramof
diagram ofoptimized
of optimizedCNN
optimized CNNfeature
featureextraction
extractioneffects.
effects.
Basedon
Based onthe
theabove
aboveanalysis,
analysis,this
thispaper
paperdecided
decidedto
touse
usethe
thewood
woodimage
imageprocessing
processingframe
frameshown
shown
in Figure 4 to obtain the wood defect feature map.
in Figure 4 to obtain the wood defect feature map.
Inputwood
Input wood
image
image
NSST
NSST
processing
processing
Feature
Feature
Extraction
Extraction
Input
Input
trainingdata
training data
Initial
Initial Superpixel
Super pixel Input
Input
trainingCNN
training CNN clusteranalysis
cluster analysis testingdata
testing data
Superpixel
Super pixelenhanced
enhanced
CNNmodel
CNN model
FeatureImage
Feature Image
Figure 4.4. Wood
Figure4. image
Woodimage processing
processingframe.
imageprocessing frame.
Figure Wood frame.
2.3. Extreme Learning Machine (ELM)
2.3.Extreme
2.3. ExtremeLearning
LearningMachine
Machine(ELM)(ELM)
Depth learning is commonly used in target recognition and defect detection applications due to
Depthlearning
Depth learningisiscommonly
commonlyused usedinintarget
targetrecognition
recognitionand anddefect
defectdetection
detectionapplications
applicationsdue dueto to
its excellent feature extraction capability. The CNN is highly time-consuming due to the necessity of
its
its excellent
excellent feature extraction
feature extraction capability.
capability. The
The CNN is highly time-consuming due to the necessity of
iterative pre-training and fine-tuning stages; theCNN is highly
hardware time-consuming
requirements for more due to the necessity
complex engineering of
iterative pre-training
iterative pre-training and and fine-tuning
fine-tuning stages;
stages; the the hardware
hardware requirements
requirements for for moremore complex
complex
applications are also high. The deep CNN structure has a large quantity of adjustable free parameters,
engineering
engineering applications are also high. The deep CNN structure has a large quantity of adjustable
which makesapplications
its constructionare also
highlyhigh. The deep
flexible. On the CNN otherstructure
hand, ithas a large
lacks quantity
theoretical of adjustable
guidance and is
freeparameters,
free parameters,which whichmakes
makesits itsconstruction
constructionhighly
highlyflexible.
flexible.On Onthetheother
otherhand,
hand,ititlacks
lackstheoretical
theoretical
overly reliant on experience, so its generalization performance is dubious. In this study, we integrated
guidance and
guidance and isis overly
overly reliant
reliant onon experience,
experience, so so its
its generalization
generalization performance
performance isis dubious.
dubious. In In this
this
the ELM into a depth extreme learning machine (Figure 5) to improve the training efficiency of the deep
study,
study, we integrated
we integrated the ELM into a depth extreme learning machine (Figure 5) to improve the
convolution network. theThe ELM into method
proposed a depth extracts
extremewood learning
defectsmachine
by using (Figure 5) to improve
an optimized CNN and the
trainingefficiency
training efficiencyof ofthe
thedeep
deepconvolution
convolutionnetwork.
network.The Theproposed
proposed methodextracts extractswood wooddefects
defects by
ELM classifier to exploit the excellent feature extraction ability of the method
deep network and fast trainingby of
using
using an
an optimized
optimized CNN
CNN and
and ELM
ELM classifier
classifier to
to exploit
exploit the
the excellent
excellent feature
feature extraction
extraction ability
ability ofthe
of the
ELM simultaneously.
deepnetwork
deep network andfast fasttraining
trainingof ofELM
ELMsimultaneously.
simultaneously.
The ELMand algorithm differs from traditional pre-feedback neural network training learning.
TheELM
The ELMalgorithm
algorithmdiffers
differsfrom
fromtraditional
traditionalpre-feedback
pre-feedback neuralnetwork networktraining
traininglearning.
learning.Its Its
Its hidden layer does not need to be iterated, and input weightsneural and hidden layer node biases are set
hidden
hidden layer
layer does not
does not needneed to
to bebe iterated,
iterated, and input
and input weights
weights and
and hidden
hidden layer
layer node biases are set
randomly to minimize training error. The output weights of the hidden layer arenode biases are
determined set
by the
randomlyto
randomly tominimize
minimizetraining
trainingerror.
error.The
Theoutput
outputweights
weightsof ofthe
thehidden
hiddenlayerlayerarearedetermined
determinedby bythethe
algorithm [21–23]. The ultimate learning machine is based on the proved ordinary extreme theorem
algorithm[21–23].
algorithm [21–23].TheTheultimate
ultimatelearning
learningmachine
machineisisbasedbasedon onthetheproved
provedordinary
ordinaryextreme
extreme theorem
and interpolation theorem, under which when the hidden layer activation function of a singletheorem
hidden
and
and interpolation
interpolation theorem,
theorem, under
under which
which when
when the
the hidden
hidden layer
layer activation
activation function
function of
of aasingle
singlehidden
hidden
layer feedforward neural network is infinitely differentiable, its learning ability is independent of the
layerfeedforward
layer feedforwardneural neuralnetwork
networkisisinfinitely
infinitelydifferentiable,
differentiable,its itslearning
learningability
abilityisisindependent
independentof ofthe
the
hiddenlayer
hidden layerparameters
parametersand andisisonly
onlyrelated
relatedtotothe
thecurrent
currentnetwork
networkstructure.
structure.WhenWhenthe theinput
inputweights
weightsAppl. Sci. 2020, 10, 7488 6 of 14
Appl. Sci. 2020, 10, x FOR PEER REVIEW 6 of 14
hidden layer parameters and is only related to the current network structure. When the input weights
and
andhidden
hidden layer
layer node
node offsets
offsetsare
arerandomly
randomlyassigned
assigned to
to obtain
obtain the
the appropriate
appropriate network
network structure,
structure,
the
theELM
ELMhas hasuniversal
universalapproximation
approximationcapability.
capability.The
Thenetwork
networkinput
inputweights
weightsandandhidden
hiddenlayer
layernode
node
offsets
offsets can be randomly assigned by approximating any continuous function. Under the premiseof
can be randomly assigned by approximating any continuous function. Under the premise of
network
networkhidden
hiddenlayer
layeractivation
activationfunction
functioninfinite
infinitedifferentiability,
differentiability,the
theoutput
outputweights
weightsof ofthe
thenetwork
network
can
canbebecalculated
calculatedvia
viathe
theleast
leastsquare
squaremethod.
method.The
Thenetwork
networkmodel
modelthat
thatcan
canapproximate
approximatethe thefunction
function
can
can be established, and the corresponding neural network functions such as classification, regression,
be established, and the corresponding neural network functions such as classification, regression,
and
andfitting
fittingcan
canbe
berealized.
realized.
Figure5.5. Contrast
Figure Contrastdiagram
diagramof
ofoptimized
optimizedCNN
CNNfeature
featureextraction
extractioneffects.
effects.
This
Thispaper
papermainly
mainly centers on the
centers on classification function
the classification of ELM,
function ofwhich
ELM,serveswhichtoserves
select atorelatively
select a
simple single
relatively hidden
simple layer
single neural
hidden network
layer neural asnetwork
the classifier.
as theThe traditional
classifier. neural network
The traditional neuralalgorithm
network
needs manyneeds
algorithm iterations
manyand parameters,
iterations learns slowly,
and parameters, hasslowly,
learns poor expansibility, and requires
has poor expansibility, andintensive
requires
manual
intensiveinterventions. The ELM used
manual interventions. The here
ELMrequires
used hereno iterations,
requires no theiterations,
learning speed is relatively
the learning speedfast,
is
the input weights
relatively fast, the and
inputbiases are and
weights generated
biases arerandomly,
generated therandomly,
follow-upthe does not need
follow-up to be
does notset,
needandto
relatively
be set, andfewrelatively
manual interventions
few manual are required. In
interventions aretherequired.
large sample
In thedatabase, the recognition
large sample database,rate the
of ELM is better
recognition rate than
of ELMthatisofbetter
the support
than that vector machine
of the support(SVM).
vectorFor these reasons,
machine (SVM). Forwethese
use ELMs as
reasons,
classifiers
we use ELMsto enhance recognition
as classifiers efficiency
to enhance and performance
recognition [24,25].
efficiency and performance [24,25].
The
TheELM
ELMalgorithm
algorithmintroduced
introducedabove aboveisisthethemain
mainclassification
classificationmethod
methodfor forwood
wooddefect
defectfeature
feature
recognition
recognitionininthis paper.
this paper. However,
However, in the inELM
the network structure,
ELM network the inputthe
structure, weights
input and the threshold
weights and the
of hidden layer
threshold nodes layer
of hidden are given
nodesrandomly.
are given Forrandomly.
the ELM structure
For the ELMwith the same number
structure with theofsame
hidden layer
number
neurons,
of hidden thelayer
performance
neurons, of the network
performance is veryofdifferent,
the networkwhich ismakes
verythe classification
different, which performance
makes the
of the networkperformance
classification more unstable. of The
the genetic
networkalgorithm (GA) simulates
more unstable. Darwinian
The genetic evolutionary
algorithm theory
(GA) simulates
to optimize the initial weights and threshold of ELM by eliminating less-fit
Darwinian evolutionary theory to optimize the initial weights and threshold of ELM by eliminating weights and thresholds.
Figure
less-fit 6a,c and
weights show the variation curves of GA-ELM population fitness functions under Radbas,
thresholds.
Hardlim, and6a,c
Figure Sigmoid
showexcitation functions,
the variation curves respectively.
of GA-ELM Smaller fitnessfitness
population values functions
indicate higher
underaccuracy;
Radbas,
the Sigmoidand
Hardlim, incentive
Sigmoid function has the
excitation best network
functions, effect. Smaller fitness values indicate higher
respectively.
accuracy; the Sigmoid incentive function has the best network effect.Appl. Sci. 2020, 10, x FOR PEER REVIEW 7 of 14
Appl. Sci. 2020, 10, 7488 7 of 14
(a)
(b)
(c)
Figure6.6. Variation curves
Figure curves ofof driving
driving function
functionpopulation
populationfitness;
fitness;(a)(a)Radbas
Radbasdriving function;
driving (b)
function;
Hardlim
(b) Hardlimdriving function;
driving (c)(c)
function; Sigmoid driving
Sigmoid function.
driving function.Appl. Sci. 2020, 10, 7488 8 of 14
The classification accuracy of GA-ELM and ELM under different excitation functions is also shown
in Table 1. We found that the classification accuracy of Sigmoid and Radbas excitations were similar,
and the Hardlim excitation function was an exception. The accuracy of ELM and GA-ELM was highest
when the Sigmoid function was used as the activation function. The accuracy of GA-ELM reached
95.93%, which is markedly better than that of an unoptimized ELM. The GA optimized ELM network
required fewer hidden layer nodes and showed higher test accuracy as well.
Table 1. Classification accuracy of GA-ELM and ELM under different excitation functions.
Excitation Classification Accuracy Rate (%) Number of Hidden Node
Function GA-ELM ELM GA-ELM ELM
Sigmoid 95.93 93.85 70 90
Hardlim 93.98 91.37 90 170
Radbas 95.37 93.36 110 200
In summary, an improved depth extreme learning machine was constructed in this study by
combining the optimized GA-ELM classifier with the optimized CNN feature extraction. It is referred
to from here on as “D-ELM”.
3. Experimentation
3.1. Experimental Parameters
Table 2 shows the computer-related parameters and software platform used by the experimental
system, including CPU model, main frequency and memory size.
Table 2. Experimental parameters.
Items Type
Server Intel(R)Xeon(R)CPU E5-4603v2
Quantity of physical CPUs 2
Main frequency of CPUs 2.20 GHz
Memory 16 GB
Anaconda 3.5,
Experimental platform
Microsoft Visual Studio 2017
3.2. Empirical Method and Result
The specific experimental process is shown in Figure 7. First, we preprocessed 5000 original
images via NSST and randomly selected 4000 images for training. Second, for each pixel in each
image, the neighborhood subgraph was taken as the input of CNN, and a total of 40,000,000 samples
were obtained as the experimental training set to train the network model. The remaining 10,000,000
samples were used as test images to evaluate the algorithm. The features extracted from the test
samples were input into the ELM network classifier, the number of hidden layer nodes of the extreme
learning machine was set to 100, then the accuracy and stability of the feature extraction method was
statistically analyzed. We found that when the number of iterations exceeds 3500, the loss function is
around 0.2 and the convergence performance is acceptable.Appl. Sci. 2020, 10, x FOR PEER REVIEW 9 of 14
Appl. Sci. 2020, 10, 7488 9 of 14
Appl. Sci. 2020, 10, x FOR PEER REVIEW 9 of 14
Figure 7. Flowchart of wood defect feature extraction and classification process.
Figure 7. Flowchart of wood defect feature extraction and classification process.
Figure 8a Figure
shows7.the
Flowchart of wood defect feature extraction and classification process.
relationship between the training loss value and the number of iterations.
Although the training loss value fluctuates
Figure 8a shows the relationship a little
between theduring theloss
training iteration
valueprocess,
and the itnumber
shows aofdownward
iterations.
Figure 8a shows the relationship between the training loss value and the number of iterations.
trend as athe
Although whole. When
training the
loss iteration
value is completed,
fluctuates the training
a little during loss value
the iteration was itaround
process, shows 0.2. Figure 8b
a downward
Although the training loss value fluctuates a little during the iteration process, it shows a downward
showsasa agraph
trend whole. ofWhen
accuracy. When the
the iteration is number of iterations
completed, wasloss
the training 1500, thewas
value accuracy
around of 0.2.
the Figure
proposed
8b
trend as a whole. When the iteration is completed, the training loss value was around 0.2. Figure 8b
algorithm
shows reached
a graph 90%. Accuracy
of accuracy. Whencontinued
the numberto increase as iteration
of iterations was 1500,quantity increased
the accuracy of until reaching
the proposed
shows a graph of accuracy. When the number of iterations was 1500, the accuracy of the proposed
a maximum
algorithm of about
reached 98%.
90%. Accuracy continued to increase as iteration quantity increased until reaching a
algorithm reached 90%. Accuracy continued to increase as iteration quantity increased until reaching
maximum of about 98%.
a maximum of about 98%.
(a)
(a) 8. Cont.
FigureAppl.Sci.
Appl. Sci.2020,
2020,10,
10,xxFOR
FORPEER
7488 PEERREVIEW
REVIEW 10ofof14
10 14
(b)
(b)
Figure8.
Figure 8.8.Relationship
Relationshipbetween
betweenlosslossfunction,
function,accuracy,
accuracy,iteration
accuracy, iterationnumber:
iteration number:(a)
number: (a)Relationship
Relationshipbetween
Relationship between
loss
lossfunction
functionand
anditeration;
iteration;(b)
(b)accuracy
accuracygraph.
graph.
graph.
Figure
Figure999shows
Figure shows
shows our
ourfinal
our recognition
final
final effect
recognition
recognition on the
effect
effect ontest
on theset.
the testWe
test surround
set.
set. the identified
We surround
We surround wood defects
the identified
the identified wood
wood
with different
defectswith
defects colored
withdifferent rectangular
differentcolored borders
coloredrectangular
rectangularborders
borders
Figure9.9.Recognition
Figure Recognitionresults
resultsbased
basedon
onbased
basedon
ondeep
deeplearning.
learning.
4. Discussion
4.4.Discussion
Discussion
This
This paper
This paperproposes
paper proposesan
proposes anELM
an ELMclassifier
ELM classifierbased
classifier basedon
based ondepth
on depthstructure.
depth structure. Choosing
structure. Choosing the
Choosing theappropriate
the appropriate
appropriate
number
numberof
number of hidden
ofhidden nodes
hiddennodes under
nodesunder the
underthe D-ELM
theD-ELM structure
D-ELMstructure provides
structureprovides enhanced
providesenhanced stability
enhancedstability and
stabilityand generalization
andgeneralization
generalization
ability
abilityin
ability inthe
in thenetwork.
the network.To
network. To
Toensure
ensureaccurate
ensure tests
accuratetests
accurate and
testsand prevent
andprevent node
preventnode redundancy,
redundancy,when
noderedundancy, whenthe
when thenumber
the numberof
number of
of
hidden
hidden nodes
nodes was
was 100,
100,the
thetest
testaccuracy
accuracy of
of D-ELM
D-ELM was
was maintained
maintained at
ata arelatively
relatively
hidden nodes was 100, the test accuracy of D-ELM was maintained at a relatively stable value over stable
stable value
valueover
over
repeated
repeatedtests.
repeated The
tests.The
tests. accuracy
Theaccuracy was
accuracywaswasphased
phasedas
phased asshown
as shownin
shown inFigure
in Figure10.
Figure 10.D-ELM
10. D-ELMsignificantly
D-ELM significantlyoutperformed
significantly outperformed
outperformed
ELM
ELM with small fluctuations in amplitude, robustness to the number of test iterations, and
ELM with
with small
small fluctuations
fluctuations in
in amplitude,
amplitude, robustness
robustness to
tothe
thenumber
number of
of test
testiterations,
iterations, andhigher
and higher
higher
network stability.
networkstability.
network stability.Appl. Sci. 2020, 10, 7488 11 of 14
Appl. Sci. 2020, 10, x FOR PEER REVIEW 11 of 14
Figure 10.Accuracy
Figure10. AccuracyofofD-ELM
D-ELMand
andELM
ELMafter
aftermultiple
multipletests.
tests.
Table
Table3 shows thethe
3 shows results of our
results of algorithm accuracy
our algorithm tests, as
accuracy mentioned
tests, above. above.
as mentioned D-ELMD-ELM
has a higher
has a
average accuracyaccuracy
higher average rate but lower
rate butstandard deviationdeviation
lower standard than ELM. TheELM.
than accuracy and stability
The accuracy and of D-ELMof
stability
network
D-ELM were bothwere
network accurate
bothand stable.and
accurate As astable.
result,As
theaperformance of the classifier
result, the performance wasclassifier
of the improved. was
improved.
Table 3. D-ELM versus ELM stability.
Table 3. D-ELM versus ELM stability.
Average Accuracy Highest Accuracy Lowest Accuracy Standard
Algorithms
Rate (%) Rate (%) Rate (%) Deviation (std)
Average Accuracy Highest Accuracy Lowest Accuracy Standard
Algorithms
D-ELM 95.92 96.11 95.65 0.0967 (std)
Rate (%) Rate (%) Rate (%) Deviation
ELM 92.84 93.72 92.16 0.3220
D-ELM 95.92 96.11 95.65 0.0967
ELM 92.84 93.72 92.16 0.3220
We added an SVM classifier to the experiment to further assess the depth extreme learning
machine.
We Table
added4 anshows
SVMthe accuracy
classifier to and
the timing of D-ELM
experiment and SVM
to further assesstraining testsextreme
the depth on all samples,
learning
where D-ELM
machine. again
Table has the
4 shows thehighest accuracy
accuracy in bothoftraining
and timing D-ELMand andtesting. Although
SVM training theon
tests training time
all samples,
and network
where D-ELM layer quantity
again has theare higher
highest in D-ELM,
accuracy its training
in both trainingtime and testAlthough
and testing. time are shorter than time
the training the
other algorithms
and network wequantity
layer tested, and
are its accuracy
higher is much
in D-ELM, its higher.
trainingThe
timeoverall performance
and test of D-ELM
time are shorter is
than the
otherthan
better algorithms we tested,
that of ELM and its accuracy is much higher. The overall performance of D-ELM is
and SVM.
better than that of ELM and SVM.
Table 4. Defect recognition of D-ELM ELM and SVM on wood images.
Table 4. Defect recognition of D-ELM
Algorithm ELM and
D-ELM SVM on wood
ELM SVMimages.
Algorithm
Time (s) D-ELM 1879
890 ELM SVM
2356
Train
Accuracy rate (s)
Time (%) 99.47%890 92.31%
1879 90.45%
2356
Train
Accuracy
Time (ms)rate (%) 187
99.47% 532
92.31% 90.45%
631
Test
Accuracy rate (%)
Time (ms) 96.72187 92.16
532 91.55
631
Test
Accuracy rate (%) 96.72 92.16 91.55
5. System Interface
5. System Interface
We constructed a network model and classification optimizer based on the proposed algorithm by
integrating Anaconda 3.5
We constructed and TensorFlow.
a network model and We then constructed
classification a realbased
optimizer woodon plate
thedefect identification
proposed algorithm
system in the C# development
by integrating Anaconda 3.5 language on the Microsoft
and TensorFlow. We Visual Studio 2017 open
then constructed platform.
a real The system
wood plate defect
can identify defects
identification in solid
system wood
in the C#plate images and
development provide their
language position
on the and size
Microsoft information.
Visual Studio 2017We also
open
used a Microsoft
platform. SQL server
The system 2012 database
can identify defects into solid
store wood
the information before
plate images andand after their
provide processing.
position and
sizeOur
information.
experiment Weonalso used
deep a Microsoft
network SQL
feature server 2012
learning mainlydatabase
involvedto store the informationof
the implementation before
the
and after
network processing.
framework and the training model for wood recognition. The system is based on the network
trainingOur experiment
model discussed onindeep network
this paper; feature
it can learning
be used mainly
to detect involved
defects the implementation
in the wood image databaseof onthe
a
network
single sheetframework
and displayandthe the traininginmodel
coordinates for Y
the X and wood recognition.
directions The system
of the defects, as shownis based on 11.
in Figure the
Onnetwork
the lefttraining
side, themodel
scanned discussed in thisare
wood images paper; it can with
displayed be used to detect
defects markeddefects in the
in green wood
boxes. Theimage
top
database on a single sheet and display the coordinates in the X and Y directions of the defects, as
shown in Figure 11. On the left side, the scanned wood images are displayed with defects marked inAppl. Sci. 2020, 10, 7488 12 of 14
Appl. Sci. 2020, 10, x FOR PEER REVIEW 12 of 14
right side of the interface shows the coordinates of each defect, the cutting position of the plank, and
green boxes. The top right side of the interface shows the coordinates of each defect, the cutting
the type of defects. Below the table are the total numbers of defects identified by the machine and the
position of the plank, and the type of defects. Below the table are the total numbers of defects
recognition rate.
identified by the machine and the recognition rate.
Figure
Figure 11.11. Systeminterface
System interface diagram
diagram based
basedonondepth
depthlearning.
learning.
6. Conclusions
6. Conclusions
The depth
The depth extremeextreme learning
learning machine machine
proposed proposed
in thisinpaper
this paper has reasonable
has reasonable dimensions,
dimensions, effectively
effectively manages heterogeneous data, and works within an acceptable run time. Our results
manages heterogeneous data, and works within an acceptable run time. Our results suggest that it
suggest that it is a promising new solution to problems such as obtaining marking samples,
is a promising new solution to problems such as obtaining marking samples, constructing features,
constructing features, and training. It has excellent feature extraction ability and fast training time.
and training.
Based on It thehas excellent
method featurelearning,
of machine extraction Theability and fast training
NSST transform is used totime. Basedthe
preprocess on original
the method
of machine learning,
image (i.e., reduce itsThe NSST transform
complexity is used to
and dimensionality preprocess
while minimizing thetheoriginal image (i.e.,
down-sampling processreduce
its complexity
in CNN), then andSLICdimensionality while minimizing
is applied to optimize the CNN model thetraining
down-sampling
process. This process
methodin CNN), then
effectively
SLICreduces
is applied to optimizeofthe
the redundancy CNN
local imagemodel training
information andprocess.
extracts This method
relatively effectively
accurate reduces the
supplementary
feature contours. The optimized CNN is then used to extract wood
redundancy of local image information and extracts relatively accurate supplementary feature image features and secure
contours.
corresponding image features. The feature is input to the ELM
The optimized CNN is then used to extract wood image features and secure corresponding classifier and the parameters of theimage
related
features. Theneural network
feature are optimized.
is input to the ELM The GA is used
classifier andtotheselect the initial weight
parameters of the threshold of ELM
related neural to
network
improve the prediction accuracy and stability of the network model. Finally, the image data to be
are optimized. The GA is used to select the initial weight threshold of ELM to improve the prediction
tested is input to the well-trained network model and final test results are obtained.
accuracy and stability of the network model. Finally, the image data to be tested is input to the
We also compared the stability of D-ELM and ELM network models. The standard deviation of
well-trained
D-ELM was network modeland
only 0.0967 andthefinal test results
accuracy of D-ELMare obtained.
improved by about 3% compared to ELM; the
We
stability of the D-ELM network was also found to beELM
also compared the stability of D-ELM and highernetwork
and lessmodels. Thetest
affected by standard
quantitydeviation
than
of D-ELM was only 0.0967 and the accuracy of D-ELM improved by
ELM. We also found that D-ELM has an accuracy of up to 96.72% and a shorter test time than ELMabout 3% compared to ELM;
the stability
or SVMofatthe onlyD-ELM
187 ms. network
The D-ELMwas also foundmodel
network to be is higher
capableandofless affected
highly by test
accurate woodquantity
defect than
ELM.recognition
We also foundwithinthat
a very short has
D-ELM training and detection
an accuracy of uptime.
to 96.72% and a shorter test time than ELM or
SVM Author
at onlyContributions:
187 ms. The Conceptualization,
D-ELM networkY.Y. model and is capable
Y.L.; of highly
methodology, Y.Y.;accurate
software,wood defect recognition
X.Z.; validation, Y.Y.,
within a very
Z.H. short
and F.D.; training
resources, and
X.Z. anddetection time.
Z.H.; writing—original draft preparation, Y.Y.; writing—review and editing,
Y.L. and X.Z.; project administration, Y.L.; funding acquisition, Y.L. All authors have read and agreed to the
Author version ofConceptualization,
Contributions:
published the manuscript. Y.Y. and Y.L.; methodology, Y.Y.; software, X.Z.; validation, Y.Y., Z.H.
and F.D.; resources, X.Z. and Z.H.; writing—original draft preparation, Y.Y.; writing—review and editing, Y.L. and
Funding:
X.Z.; project This research was
administration, funded
Y.L.; by the
funding 2019 JiangsuY.L.
acquisition, Province Key Research
All authors and Development
have read Plan
and agreed to theby the
published
Jiangsu
version of theProvince Science and Technology under grant BE2019112, and was funded by the Jiangsu Province
manuscript.
International Science and Technology Cooperation Project under grant BZ2016028, and was supported from the
Funding: This research was funded by the 2019 Jiangsu Province Key Research and Development Plan by the
Jiangsu Province Science and Technology under grant BE2019112, and was funded by the Jiangsu Province
International Science and Technology Cooperation Project under grant BZ2016028, and was supported from the
948 Import Program on the Internationally Advanced Forestry Science and Technology by the State Forestry
Bureau under grant 2014-4-48.Appl. Sci. 2020, 10, 7488 13 of 14
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Qiu, Q.W.; Lau, D. Grain Effect on the Accuracy of Defect Detection in Wood Structure by Using Acoustic-Laser
Technique. In Proceedings of the Nondestructive Characterization and Monitoring of Advanced Materials,
Aerospace, Civil Infrastructure, and Transportation XIII, Denver, CO, USA, 3–7 March 2019. [CrossRef]
2. Siekanski, P.; Magda, K.; Malowany, K.; Rutkiewicz, J.; Styk, A.; Krzeslowski, J.; Kowaluk, T.; Zagorski, A.
On-line laser triangulation scanner for wood logs surface geometry measurement. Sensors 2019, 19, 1074.
[CrossRef] [PubMed]
3. Espinosa, L.; Prieto, F.; Brancheriau, L.; Lasaygues, P. Effect of wood anisotropy in ultrasonic wave
propagation: A ray-tracing approach. Ultrasonics 2019, 91, 242–251. [CrossRef] [PubMed]
4. Taskhiri, M.S.; Hafezi, M.H.; Harle, R.; Williams, D.; Kundu, T.; Turner, P. Ultrasonic and thermal testing
to non-destructively identify internal defects in plantation eucalypts. Comput. Electron. Agric. 2020, 173.
[CrossRef]
5. Yang, H.M.; Yu, L. Feature extraction of wood-hole defects using wavelet-based ultrasonic testing. J. For. Res.
2017, 28, 395–402. [CrossRef]
6. Lukomski, M.; Strojecki, M.; Pretzel, B.; Blades, N.; Beltran, V.L.; Freeman, A. Acoustic emission monitoring
of micro-damage in wooden art objects to assess climate management strategies. Insight 2017, 59, 256–264.
[CrossRef]
7. Rescalvo, F.J.; Valverde-Palacios, I.; Suarez, E.; Roldan, A.; Gallego, A. Monitoring of carbon fiber-reinforced
old timber beams via strain and multiresonant acoustic emission sensors. Sensors 2018, 18, 1224. [CrossRef]
[PubMed]
8. Li, C.; Zhang, Y.; Tu, W.; Jun, C.; Liang, H.; Yu, H. Soft measurement of wood defects based on LDA feature
fusion and compressed sensor images. J. For. Res. 2017, 28, 1285–1292. [CrossRef]
9. Dai, J.; Li, Y.; He, K.; Sun, J. R-FCN: Object detection via region-based fully convolutional networks.
arXiv 2016, arXiv:1605.06409.
10. Szegedy, C.; Liu, W.; Jia, Y.Q.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; Rabinovich, A.
Going Deeper with Convolutions. In Proceedings of the 2015 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), Boston, MA, USA, 7–12 June 2015; pp. 1–9. [CrossRef]
11. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [CrossRef]
12. He, T.; Liu, Y.; Yu, Y.B.; Zhao, Q.; Hu, Z.K. Application of deep convolutional neural network on feature
extraction and detection of wood defects. Measurement 2020, 152. [CrossRef]
13. Hu, K.; Wang, B.J.; Shen, Y.; Guan, J.R.; Cai, Y. Defect identification method for poplar veneer based
on progressive growing generated adversarial network and MASK R-CNN Model. Bioresources 2020, 15,
3041–3052. [CrossRef]
14. Shi, J.H.; Li, Z.Y.; Zhu, T.T.; Wang, D.Y.; Ni, C. Defect detection of industry wood veneer based on NAS and
multi-channel mask R-CNN. Sensors 2020, 20, 4398. [CrossRef] [PubMed]
15. Yang, W.X.; Jin, L.W.; Tao, D.C.; Xie, Z.C.; Feng, Z.Y. DropSample: A new training method to enhance deep
convolutional neural networks for large-scale unconstrained handwritten Chinese character recognition.
Pattern Recogn. 2016, 58, 190–203. [CrossRef]
16. Wan, W.; Lee, H.J. Multi-focus image fusion based on non-subsampled shearlet transform and sparse
representation. Lect. Notes Electr. Eng. 2018, 449, 120–126. [CrossRef]
17. Singh, S.; Anand, R.S.; Gupta, D. CT and MR image information fusion scheme using a cascaded framework
in ripplet and NSST domain. IET Image Process 2018, 12, 696–707. [CrossRef]
18. Wu, W.; Qiu, Z.M.; Zhao, M.; Huang, Q.H.; Lei, Y. Visible and infrared image fusion using NSST and deep
Boltzmann machine. Optik 2018, 157, 334–342. [CrossRef]
19. Boemer, F.; Ratner, E.; Lendasse, A. Parameter-free image segmentation with SLIC. Neurocomputing 2018, 277,
228–236. [CrossRef]
20. Chen, L.C.; Papandreou, G.; Kokkinos, I.; Murphy, K.; Yuille, A.L. DeepLab: Semantic image segmentation
with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE Trans. Pattern Anal.
Mach. Intell. 2018, 40, 834–848. [CrossRef]Appl. Sci. 2020, 10, 7488 14 of 14
21. Zhu, W.T.; Miao, J.; Qing, L.Y.; Huang, G.B. Hierarchical Extreme Learning Machine for Unsupervised
Representation Learning. In Proceedings of the Joint Conference on Neural Networks, Killarney, Ireland,
11–16 July 2015.
22. Uzair, M.; Shafait, F.; Ghanem, B.; Mian, A. Representation learning with deep extreme learning machines
for efficient image set classification. Neural Comput. Appl. 2018, 30, 1211–1223. [CrossRef]
23. Tang, J.X.; Deng, C.W.; Huang, G.B. Extreme learning machine for multilayer perceptron. IEEE Trans. Neural
Networks Learn. Syst. 2016, 27, 809–821. [CrossRef]
24. Yu, W.C.; Zhuang, F.Z.; He, Q.; Shi, Z.Z. Learning deep representations via extreme learning machines.
Neurocomputing 2015, 149, 308–315. [CrossRef]
25. Liu, S.L.; Feng, L.; Xiao, Y.; Wang, H.B. Robust activation function and its application: Semi-supervised
kernel extreme learning method. Neurocomputing 2014, 144, 318–328. [CrossRef]
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional
affiliations.
© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).You can also read
