1. Robust Image Hashing Based on Color Vector Angle and Canny Operator
Source: AEU-International Journal of Electronics and Communications (2016)
Authors: Zhenjun Tang, Liyan Huang, Xianquan Zhang, Huan Lao
Speaker: Meihui Sun
Date:

2. Outline Introduction Proposed Scheme Experimental Results Conclusions
- Color vector angle calculation
- Edge detection - Statistical feature extraction
Experimental Results
- Perceptual robustness
- Discriminative capability
- Effect of circle number
- Performance comparisons
- Reference
Conclusions

3. Introduction Application: image forensics, retrieval, index
Convention image hashing algorithms have limitation in robustness and discrimination.
This hashing robust against normal digital processing.

4. Proposed Scheme(1/9) - Frame
This image hashing consists of three steps.

5. Proposed scheme(2/9) - Preprocessing
Original Image
512*512 Image
Processed Image
resized
blurred
The convolution mask of Gaussian low-pass filter:
Where is the standard deviation of all elements in the convolution mask.
卷积模板中所有元素的标准差

6. Proposed scheme(3/9) - Calculation of color vector angle
Luminance
Features of a color image: Hue
Saturation
Comparing with RGB space:

7. Proposed scheme(4/9) - Calculation of color vector angle
The color vector angle θ can be calculated by:
To reduce computational cost:

8. Proposed scheme(5/9) - Calculation of color vector angle
Reference color Pref = [Rref， Gref， Bref] :
Ri,j, Gi,j and Bi,j are the red, green and blue components of the image pixel Pi,j in
the i-th row and the j-th column (1 ≤ i ≤ M, 1 ≤ j ≤ M).

9. Proposed scheme(6/9) - Calculation of color vector angle
Matrix Acolor(the color vector angle sinθi,j between Pi,j and Pref ):
Conversion from the color image to color vector angle:

10. Proposed scheme(7/9) - Edge detection
Steps:
1.Convert the preprocessed image to YCbCr color space;
2.Take the luminance component for representation;
3.Apply the Canny operator to the luminance component.
1, Pi,j is an edge
point.
Ei,j=
0, Pi,j is not an
edge point.

11. Proposed scheme(8/9) - Statistical feature extraction
K the number of circles.
rk the k-th radius labeled (r1and rk are the innermost and outermost radii).
di,j the distance from the pixel Pi,j(1≤i≤M, 1≤j≤M) to the image center (xc,yc). d M

12. Proposed scheme(9/9) - Statistical feature extraction
E(k)--- the set of color vector angle of those edge pixels on the k-th concentric circle
Δd : a pre-defined threshold for error control.
Nk : the element number of E(k)
Quantized:
Hash h:
Consequently, E(k) can be determined by the following equation:
The variance vk:

13. Experimental Results (1/8) - Similarity evaluation
u1 and u2 : the means of h1 and h2.
h1(k) and h2(k) : the k-th elements of h1 and h2.
A threshold T:
S＞T similar images
S≤T different image
Use the correlation coefficient to measure similarity between two image hashes

14. Experimental Results (2/8) - Perceptual robustness
42 color images:
Including Airplane, Baboon, House, Peppers, and Lena and
37 images (all color images) of the USC-SIPI Image Database.

15. Experimental Results (3/8) - Perceptual robustness

16. Experimental Results (4/8) - Discriminative capability
200 different color image:

17. Experimental Results (5/8) - Discriminative capability
200*199/2=19000.
Minimum:
Maximum:
Mean:
Standard deviation:
There are, pairs of different images
If T=0.95,there are only 0.04% different image wrongly considered as similar images.

18. Experimental Results (6/8) - Effect of circle number
Receiver operating characteristics (ROC):
True positive rate :
False positive rate:
K=40

19. Experimental Results (7/8) - Performance comparisons

20. Experimental Results (8/8) - Performance comparisons
[1]Tang Z, Dai Y, Zhang X, Zhang S. Perceptual image hashing with histogram of color vector angles. In: The 8th international conference on active media technology
(AMT 2012) p. 237–46.
[2]Li Y, Lu Z, Zhu C, Niu X. Robust image hashing based on random Gabor filtering and dithered lattice vector quantization. IEEE Trans Image Process.
[3]Ou Y, Rhee KH. A key-dependent secure image hashing scheme by using Radon transform. In: Proceedings of the IEEE international symposium on Intelligent signal processing and communication systems p. 595–8.
[4]Laradji IH, Ghouti L, Khiari E-H. Perceptual hashing of color images using hyper-complex representations. In: Proceedings of the IEEE international conference on image processing (ICIP 2013) p. 4402–6.

21. Conclusions
1.This HASH is robust against normal content-preserving manipulations.
2.This hash has a good performances in robustness and discriminative capability.

22. Thanks!