Animal identification using machine learning techniques
1. 1
Animal Identification using Machine Learning
Techniques
By
Aya Salama Abdelhady
Under the Supervision of:
Professor Aly Fahmy & Professor Abou Elella Hassanein
Ph.D Presentation
Department of Computer Science
2. Agenda
Introduction
Motivation
Problem Statement
Research Objectives
Literature Review and Current Approaches
Proposed Approach (Deep Neural Networks)
Data Sets
Work plan
2
3. 3
Introduction
Animal identification refers to the
recognition process of animals.
Classical animal identification methods
such as ear tags and tattooing are limited
for decision support due to their
vulnerability to loss and manipulation.
Biometric mapped into animal
identification systems are a promising
trend owing to their uniqueness and
immutability.
Ear notching
Branding
4. Animal identification is vital in in large group of animals
Individual identification allow management of
Stockbreeding programs
Disease and treatment
Arabian Horse identification
Arabian horses are precious and expensive
Arabian horses identification is important in international competitions
Classic methods for horses are considered as scars and also vulnerable to
manipulation
Sheep identification
Guarantee users ownership
Avoid manipulation in type and price
4
Motivation
5. Investigation of bio and physical metrics that lead to
best identification results
(Example of these features are eyes, iris, face, weight)
Data Sets Collection
Arabian horses identification has no available data
sets of horses
Sheep identification has no available data sets of
sheep
5
Problem Statement
6. 6
Thesis Objectives
The aim of the thesis are:
To develop real time mobile application for:
• Arabian horse Identification
• Sheep Identification
To build animal weight estimation module
that would help in animal identification
8. 8
Literature Review
Disadvantages of Classical Methods
Vulnerability to losses, duplications and manipulation
Hot branding and freeze branding cause a lot of pain
Difficult to read
Most of these methods are painful and can lead to
infection, and they can also be considered as scars.
9. 9
• Iris Pattern
• Muzzle Prints
• Face Pattern
• Retinal vascular
Iris is considered one of the most reliable and accurate
biometric
Literature Review
Animal Biometric Methods
10. 10
10
Literature Review
Machine Learning Approaches
Machine Learning/
Artificial Intelligence
Supervised
Learning
Unsupervised
Learning
Deep Learning Other Approaches
Regressi
on
Classificati
on
Clusterin
g
Factor
Analysis
Reinforceme
nt Learning
Semi-
supervised
Active
learning
Lasso,
Ridge,
Loess,
KNN,
Spline,
XGBoos
Logistic,
SVM,
Random
Forest,
Hidden
Markov
K-means,
Birch,
Ward
Spectral
Cluster
PCA,
ICA,
NMF
Multilayer Perceptron
(MLP), Convolutional
Neural Nets (CNN),
Long Short-Term
memory (LSTM),
Restricted Boltzman
Machine (RBM)
11. 11
Literature Review
Deep Learning
Ruggedness to shifts and distortion in the
image
Fewer memory requirements
Easier and better training
Reduces the need for feature engineering,
one of the most time-consuming parts of
machine learning practice.
It significantly outperforms other solutions in
multiple domains, this includes speech,
language, vision
Is an architecture that can be adapted to new
problems relatively easily (e.g. Vision, time
series, language)
12. 12
Literature Review
Animal identification systems
A horse identification system using biometrics System
For iris segmentation iris area was extracted by first defining rectangle around the pupil area by largest dark area recognition with
estimation the most common known radius in all collected images (17,000 digital still images)
Then labels were assigned to different part of the eye; Manual labeling was essential part in that procedure.
Then Gabor filter was used to extract iris characteristics.
For horse identification, hamming distance was calculated to measure angle difference and radius difference between different images .
FRR was reported to be 0.201 and FAR 2.55 *10^-7
Face Recognition as a Biometric Identifier of Sheep
Independent component analysis with the cosine distance classifier, has been used for sheep face recognition
Accuracy is 95.3% (has been evaluated for sheep face recognition using a small set of normalized images)
This accuracy varies under scalability
13. 13
The Suggested System
Data Preparation
Data Augmentation
Investigate deep learning architecture in generative
models
Investigate deep learning architecture in
discriminative models
14. 14
Data Sets Collections
14
Data collection was our first step to
achieve our goal very.
( 6 months of preparation)
Since there is no available
benchmark for horses' eyes, we had
to collect our own data.
Data is collected for 145 horses in
different illumination conditions and
from different angles.
1,015 images were collected
16. 161616
Corpora Nigra
During data collection, we were able to capture
the corpora nigra
Corpora nigra which lies in the iris just above
the pupil very unique shape and differs from
one horse to another .
• Pupil with corpora nigra has a color that is
significantly different from all other parts in the
images.
• K-means clustering is proposed to detect the
pupil without any need for human supervision
19. 1919
Preliminary Results and Evaluation
Iris/pupil segmentation is a main step in
the process of horses recognition.
Circular Hough transform with modified
version of Kovesi’s Canny edge detection
used to detect the circular region
containing both iris and pupil.
We could not use the same way to detect
the iris itself, because as we mentioned
before the iris of the horse is different
from the iris of human.
Circular Hough transform
20. 2020
Preliminary Results and Evaluation (2)
• Pupil with retina cogina has a color that is
significantly different from all other parts in
the images.
• K-means clustering is proposed to detect the
pupil without any need for human supervision
• Euclidean distance is calculated is used to
define the distance of the nearest centroid
• Then connected component labeling was
applied to detect the biggest blob
• Then morphological noise was applied to
remove noise
21. 2121
Preliminary Results and Evaluation (3)
• Jaccard similarity is the most common used matrix to
evaluate the performance of segmentation
• After removing all unwanted images we had total of
320 images to work on.
• Ground truth images were built for 320 images.
• For 256 images, Jaccard coefficient ranged between
80% and 95%, while for the rest 64 images ranged
between 40% and 70%
• Low coefficients were due to high brightness around
the iris and very strong reflection of other objects,
which lead to false data sometimes
Noise in eye images
24. 2424
Work Plan
1. Continue survey and identify How to best
recognize animals using deep learning.
2. Investigate generative and descriptive models
3. Continue working on data collection
4. Design and develop a model for animal
identification
5. Conduct a performance analysis of the developed
model with the existing ones. Partners in this work (Suez Canal University (agriculture faculty)
and Cairo University (Faculty Vet. medicine ))
All accepted papers (regular, short, and poster) will be published by ACM – International Conference Proceedings Series (ICPS) and will be available in ACM Digital Library . ISBN: 978-1-4503-5243-7