Undergraduate thesis
Compact Descriptors for Visual Search
for Money Recognition
Student: Pham Tran Huong Giang
Supervisor: Dr. Le Thanh Ha
1
Major: Computer Science – Faculty: Information Technology
University of Engineering and Technology
Outline
Introduction
Compact Descriptors for Visual Search
Approach
Conclusion and future work
2
Introduction
Compact Descriptors for Visual Search
International standard for visual search systems
Formed and developed by MPEG
Will be used in a great number of visual search
applications
3
Introduction
Money recognition: a practical application for businessman
and tourists. It is hard to recognize local money in the first
use when they come to a foreign country.
Motivated application: money recognition for smartphone
Study CDVS and evaluate the ability of using CDVS to solve
money recognition problem
4
Compact descriptor for visual search
Extracting compact descriptor
Retrieval
Pairwise matching
5
Compact Descriptors for Visual Search
Extracting compact descriptor
Detect key points
Select features
Aggregate global descriptor
Encode feature’s coordinate
Compact
descriptor
6
Describe
Compress local
descriptors
Local
descriptors
Compact Descriptors for Visual Search
Retrieval
Query
image
Extract
compact
descriptor
Compare
global
descriptors
Descriptor
database
7
Top
match list
Compact Descriptors for Visual Search
Pairwise matching
Referent
images
Query
image
Extract
descriptor
Match local
descriptors in
compressed
domain
Check the
geometric
consistency
Matching
8
Homography
Extract
descriptor
Approach
To evaluate:
Collect a set of images of money as training set
Build a small program to recognize money
Test by another set of images (test set)
9
Approach
Collecting training dataset
Consists of the money of 6 countries: Vietnam, Laos, Cambodia,
Japan, Thailand, Singapore
Good condition of displayed money
Without background
Uniform distributed light
238 images: 160 images of banknote and 78 images of coin
10
Approach
Building money recognition program
Input
image
Find and
extract
circle
11
Extract
descriptor
Extract
descriptor
Match with 160
descriptors of
banknote
Match with 78
descriptors of
coin
Evaluate by
compare with 2
threshold one
of banknote and
one of coin
Result
How much?
Nationality?
Approach
Testing and discussion
Test set: 6 subsets
1.
2.
3.
4.
5.
6.
12
100 images of banknotes (same type with images in training dataset,
taken by me)
50 downloaded images from the Internet (same type with training
images)
60 images of coins (same type with training images, taken by me)
25 distractor images of banknotes
10 distractor images of coins
20 distractor images without money
Testing result
For threshold of 100 for banknote and 50 for coin
Subset
Precision
Number of “not found”
Number of false results
results
13
1
70/100
30
0
2
3/60
54
3
3
46/50
3
1
4
20/20
20
0
5
18/20
18
2
6
10/10
10
0
Discussion
Some successful tests in subset 1
14
Discussion
Some “not found” tests in subset 1
Comment: Money in these images is folded or strong line shine
through the money
15
Discussion
In subset 2: 3 “not found” cases are caused by the low quality
of images
One “false” case because of the similarity between 2 images:
16
Discussion
CDVS treats well with distractor images (subset 4, 5, 6)
Only 2 wrong cases in subset 5 because the distractor images are
similar with training images
17
Discussion
For coins, the recognition accuracy is hardly acceptable
because of some reasons
Coin is poor in feature
Strongly response to light
Whole-colorized
Small surface
18
Conclusion and future work
Conclusion:
CDVS brings high recognition accuracy for banknote if the image
of banknote is taken in not too bad condition.
The recognition accuracy for coin is hardly acceptable.
Future work:
Collect more data
Build an mobile application for tourists to recognize all kind of
money in the world.
19
Thank you for watching
20