3. Machine Learning for Prediction

wonderfuldistinctAI and Robotics

Oct 16, 2013 (4 years and 28 days ago)

91 views

Strathclyde Hyperspectral Imaging
Centre

Academic: Professor and lecturer

Research:

2 post docs (sponsored by
Argans

and
TiC
)




2 PhD
students
(sponsored by Gilden and AZ)


working
on
Fundamental
research and HSI applications


HSI imaging facility donated by Gilden Photonics

About the centre

1
st

& 2
nd

UK Hyperspectral Imaging Conferences

Strathclyde
April 2010 May 2011


Sponsors/ Exhibitors
Gilden Photonics, Gooch
&
Housego
,
Specim
, Pacer,
PANDA,
Headwall
,
Photometrics
, Hamamatsu, Mapping Solutions


Academic
Partners,

D Pullan
,
Univ

Leicester,
A Harvey
, Herriot Watt ,
D Reynolds
UWE,

S Morgan,
Nottingham


3
rd

HSI Conference
INGV, Rome 15
-
16 May 2012




HSI Community building activities & partners

HSI Projects

Applications


Algae classification


Beef quality assessment


Drug dissolution process


Finger print detection


Chinese tea classification




Fundamental research


New data processing tools
-

Copulas


Band selection /data reduction methods

Quantitative Assessment of Beef Quality

with Hyperspectral Imaging

using Machine Learning Techniques

Jinchang Ren
1
, Stephen Marshall

1
,

Cameron Craigie
2,3


and Charlotte Maltin
4



1
Hyperspectral Imaging Centre, University of Strathclyde, Glasgow, U. K.

2
Scottish Agricultural College (SAC), Edinburgh, U. K.

3
IVABS, Massey University, Palmerston North, N. Z.

4
Quality Meat Scotland, Ingliston Newbridge, U. K.


Outline

1.
Introduction

2.
Feature Selection

3.
Machine Learning for Prediction

4.
Experiments Results and Discussions

5.
Conclusions



1. Introduction

NIR Hyperspectral Imaging System [14]


Hyperspectral imaging of


meat in NIR spectral ranges


To predict meat quality of


Freshness


Tenderness


Fat


Protein


Moisture content


Image processing and
machine learning techniques
like neural network are used
for prediction & classification.


Nearly non
-
invasive and
accurate

1. Introduction

Existing Work

NIR sampling over 700
-
2500nm is preferred to allow spectra
acquired in three modes, reflection, transmission and
transflection
, and also enables response to key food
components like C
-
H, O
-
H and N
-
H molecular bonds.

1)
Qiao

et al [6] used selected bands on a trained neural network
for prediction of drip loss, pH and surface colour of pork, and the
prediction correlation coefficients are 0.77, 0.56 and 0.86.

2)
Eimasry

[7] employed PCA for data reduction and partial least
square regression for prediction of surface colour, tenderness and
pH of beef. The correlations are 0.88, 0.73 and 0.81;

3)
Carigie

[8] used similar linear regression on beef data captured
under commercial conditions and the correlation is less than 35%
for surface colour, pH and
M.
longissimus

dorsi

tenderness.

2. Feature Selection


Spectral Profiles

Usually the spectral profiles are used as features yet some
spectral bands or their combinations are more distinguishable
than others thus feature selection is desirable.

2. Feature Selection

1) Spectral redundancy makes it possible for feature
selection and data reduction;

2) Band selection is not used here as though it is found to
yield good results [1, 10]. Instead, principal component
analysis (PCA) is utilised as it is a standard one which has
been integrated in many existing development tools like
Matlab
.

3) We apply PCA to the original spectral data and then
select the dominant principal components as new features
for training and prediction.

3. Machine Learning for Prediction

1) Linear regression and its variations are widely used for
prediction of relevant parameters, such as partial least
square regression (PSLR) in [7], the general linear models
(GLM) in [8] and multi
-
linear regression (MLR) in [12].

2) Since the parametric models are not necessary linear,
machine learning approaches is preferred where support
vector machine (SVM) is employed.

3) SVM enables effective regression and prediction for both
linear and non
-
linear cases and it general outperforms
neural network in machine learning.

3. Support Vector Machine (SVM)

Parameters
w

and
b

respectively refer to a weight vector and a
bias
that

can be determined in the training process through
minimizing a cost function, and
Φ

refers to a (nonlinear) mapping
to map the input vector
x

into a higher dimensional space for
easily separated by a linear hyperplane as illustrated below.

This figure illustrates the
concept of SVM where a
non
-
linearly
-
separable
problem becomes
linearly separable in the
mapped feature space.

3. SVM
-

Training

3. SVM
-

Kernel Functions & Evaluation

LibSVM is used for the
implementation of the
SVM in both training
and prediction.


The correlation is
defined to the right for
evaluation.

4. Experimental Datasets

1)
Our data is the one used in [8] captured under
commercial conditions;

2)
The data is acquired using a NIR spectrometer at
Quality Meat Scotland [15] and the spectrum ranges
from 350nm to 1800nm with a spectral resolution of
1nm;

3)
The mean and median values in the absorbance form
from 10 repetitive scans are recorded for processing.

4)
The data samples are 234 in total.

5)
A 10
-
fold cross validation scheme is used for
evaluation.

4. Experimental Results

Correlation

linear

polynomial

RBF

Number of
principal
components = 13

0.298

0.450

0.396

Number of
principal
components = 20

0.305

0.533

0.288

Table 1
: Correlation of predicted
M. longissimus dorsi

shear
force values to the ground truth under various kernel functions
and number of principal components.

4. Analysis of Results

1)
Polynomial kernel function produces the best results;
and increasing the number of principal components
helps to improve the prediction.

2)
Linear kernel function and RBF kernel function tend to
yield much worse results.
Increasing
the number of
principal components does not certainly improve the
results from RBF kernel, though slightly better results
can be found for linear kernel.

3)
This on one hand indicates that polynomial kernel
function seems the best choice for this application. On
the other hand, it explains why linear models used in [8]
generated the poor results.

4. Experimental Results

Fig. 3
. Correlation values between the ground
truth and the extracted principal components.

For some principal
components, the
correlation values
are as high as 0.2
-
0.25!


It is possible to
choose useful
principal
components for
data prediction, and
this will be
investigated in the
next stage.

4. Conclusions

1) We have applied SVM for data prediction in food
quality analysis and found polynomial kernel to
yield the best results;

2) Using PCA for dimension reduction, we have
significantly improved the prediction of M.
longissimus

dorsi

shear force correlation values
from less than 0.35 to about 0.533;

3) How to apply our approach to predict other
parameters and also to evaluate the possibility in
choosing principal components for prediction will
be investigated in the near future.

Acknowledgements

The authors would like to thank colleagues in
industry and the Scottish Agricultural College for
their help in providing and collecting the data.

Thank you for your attention!


Any Questions?