imPlag: Detecting Image Plagiarism Using Hierarchical Near Duplicate Retrieval

Paper ID: 1570186597
imPlag: Detecting Image Plagiarism Using
Hierarchical Near Duplicate Retrieval
Siddharth Srivastava, Prerana Mukherjee, Brejesh Lall
Indian Institute of Technology, Delhi
IEEE India Council
Department of Electrical Engineering
Faculty of Engineering & Technology
Jamia Millia Islamia, New Delhi, India

Introduction
The key characteristic of image plagiarism is that it may involve the
reproduction of the original image using an entirely different mode
such as hand made sketches. Image Plagiarism can be posed as a
superset of image copy detection problems.
Fig. 1. (a) Original Image (b) Plagiarised image (reproduction of the source image) (c) Copied image (considered as
strong attack by copy detection algorithms but an expected case for Image Plagiarism)
IEEE India Council

Problems ?
• Detection of similar images – Huge Databases, Interactive Time
• Plagiarism brings in innovation
• Hence involves both Research and Engineering Challenges
- Stitched from 3888 images
- One column/row pixel from each
image
So knowing your limits is
necessary
Image Courtesy:Eirik Solheim
(Image has been used for demonstrating
the extent of deformation possible in
images)
IEEE India Council

KEY CONTRIBUTIONS
Development of a hierarchical feature extraction and feature indexing
technique.
Evaluation of recent feature extraction techniques against simple,
moderate and extreme deformations.
Dataset construction for testing image plagiarism algorithms.
IEEE India Council

Dataset
• Natural Images – mountains, rivers, animals, birds etc.
• Actual scenario – too many images can be similar but might not be
plagiarized (synthetically transformed)
• So for evaluation, dataset was created since detecting image
plagiarism is not really only Content Based Image Retrieval
• Search for images on Flickr, ukbench dataset
• Find similar images using Google Reverse Image Search (Google doesn’t index
Flickr !!)
• Transformed Images – Affine, Grayscale, Color channel separation
etc. (30 transformations)
IEEE India Council

Methodology
Heirarchical
Feature
Extraction
Feature
Indexing
Search
Query
NN
Match/Exact
Matching
Ranking or
Verification
Fingerprint the image
- Perceptual Hash
- SIFT > SURF, ORB, FREAK, PCA-SIFT
Store for retrieval
- Database
- Apache Lucene
- Locality Sensitive
Hashing
Search the index
- Search LSH Index
Relevant Results ranked
at the top
- Bag of Visual Words
Histogram matching
IEEE India Council

Hierarchical Indexing
Lucene Index
Perceptual Hash
Locality Sensitive Hashing
SIFT Features
Bag of Visual Words
Images
Lucene Index Traditional
Database
IEEE India Council

Layered Retrieval
Input Image
Calculate
Perceptual Hash
of input Image
Search LSH Index
for nearest
neighbours
Rank images based
on BoVW
histogram
matching
IEEE India Council

Perceptual Hash
• Can be used for multimedia content (audio, video, images)
• Similar images have similar hash values
Scale
image to
32x32
Convert
to
grayscale
Compute
DCT
Keep first
8x8
coefficie
nts
Take
Average
(no DC
Term)
Coeff > Avg
=> 1
Coeff <= Avg
=> 0
Flatten to
64bit vector
IEEE India Council

Bag of Visual Words
• SIFT features converted to Bag of Visual Words
• More efficient than direct keypoint matching
• Observations:
• Large vocabulary size may increase false negatives
• Small vocabulary size may increase false positives
• Though there is no definite pattern on what should the vocabulary
size be
IEEE India Council

Results
• Accuracy: 81%
• Scalability
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 1000 2000 3000 4000 5000 6000 7000
Time(sec)
Number of images in the dataset
top-60 results
IEEE India Council

Conclusion
• We perform evaluations to choose best criteria and techniques for
detecting image plagiarism.
• A method is proposed, consisting of perceptual hashing and SIFT with
hierarchical approximate matching scheme.
• This scheme was able to maintain the tradeoff between time and
accuracy.
IEEE India Council

References
• E. Chalom, E. Asa, and E. Biton, “Measuring image similarity: an overview of some useful applications,”
Instrumentation & Measurement Magazine, IEEE, vol. 16, no. 1, pp. 24–28, 2013.
• C. Zauner, M. Steinebach, and E. Hermann, “Rihamark: perceptual image hash benchmarking,” in IS&T/SPIE
Electronic Imaging. International Society for Optics and Photonics, 2011, pp. 78 800X–78 800X.
• V. Voronin, V. Frantc, V. Marchuk, and K. Egiazarian, “Fast texture and structure image reconstruction using
the perceptual hash,” Image Processing: Algorithms and Systems XI, 2013.
• A. Kumar, A. Anand, A. Akella, A. Balachandran, V. Sekar, and S. Seshan, “Flexible multimedia content
retrieval using infonames,” ACM SIGCOMM Computer Communication Review, vol. 41, no. 4, pp. 455–456,
2011
• A. Gionis, P. Indyk, R. Motwani et al., “Similarity search in high dimensions via hashing,” in VLDB, vol. 99,
1999, pp. 518–529.
• V. Christlein, C. Riess, J. Jordan, and E. Angelopoulou, “An evaluation of popular copy-move forgery detection
approaches,” Information Forensics and Security, IEEE Transactions on, vol. 7, no. 6, pp. 1841– 1854, 2012.
• M. Lux and S. A. Chatzichristofis, “Lire: lucene image retrieval: an extensible java cbir library,” in Proceedings
of the 16th ACM international conference on Multimedia. ACM, 2008, pp. 1085–1088.
IEEE India Council

THANKYOU
IEEE India Council

Appendix: Dataset Images
Perceptually Similar ?
IEEE India Council

Appendix: Nature is not always greenish
IEEE India Council

Appendix: Accuracy
0
10
20
30
40
50
60
70
80
90
100
60 100 150 200 250
Accuracy(%)
Number of Results(top-N)
Accuracy(%)
IEEE India Council

Appendix: Results
Fig 2. Comparison of Feature matching techniques Fig 3. Average time taken by SIFT, SURF and Perceptual Hash
PH
SURF
SIFT
IEEE India Council

Appendix: Results
Fig 4. Comparison of ranked retrieval Fig 5. Ranked V/s Non Ranked Retrieval
IEEE India Council

Appendix: Results
Fig 6. Time vs Number of results Fig 7. Time vs Number of Images in the dataset
IEEE India Council

Appendix: Results
Fig 8. Lucene v/s Database Retrieval time
IEEE India Council

Locality Sensitive Hashing
• Similar features hashed to same hash values
• Parameters
• No of bits (k)
• No of tables (l)
• Maximum Bucket capacity (usually unlimited)
• Empirical Analysis needed for determining parameters as per the
dataset
• varying number of bits, varies bucket size (small hash, more collisions
and vice versa)
IEEE India Council

Lucene
• Very efficient in document indexing and retrieval
• Bag of Visual words histograms are indexed
• Allows for random access of documents
• Histograms are fetched from Lucene index and ranked (Filtering)
IEEE India Council

imPlag: Detecting Image Plagiarism Using Hierarchical Near Duplicate Retrieval

More Related Content

What's hot (19)

Similar to imPlag: Detecting Image Plagiarism Using Hierarchical Near Duplicate Retrieval (20)

More from Prerana Mukherjee (10)

Recently uploaded (20)

imPlag: Detecting Image Plagiarism Using Hierarchical Near Duplicate Retrieval