1. Palette Partition Based Data Hiding for Color Images Yu-Chiang Li, Piyu Tsai, Chih-Hung Lin, Hsiu-Lien Yeh, and Chien-Ting Huang Speaker : Yu-Chiang Li Date : 2009/09/13

2. 2 Outline Introduction 1 Related Work 2 Proposed Scheme 33 Experimental Results 44 Conclusions 5

3. 3 Introduction  Image data hiding employs original image to cover a secret message  Secret message is imperceptibly inserted into the original image

4. 4 Introduction (c.)  Good data hiding technologies  Imperceptibility  High embedding capacity  Trade-off between  stego-image quality  hiding capacity

5. 5 Introduction (c.)  Three categories of data hiding  Spatial domain  Frequency domain  Quantization domain

6. 6 Introduction (c.) Spatial Domain  Secret message is directly embedded into spatial domain  Least-significant-bits Cover Image 1010 0101 1100 0011 Secret Message inserted 103102103104 105108109 201 202 204 207 185187188 189191192 194 195197198 199201 5156616671768186 9095100105110115120135 161163164168171173174176 179181183185186188190192 Stego-Image 102103 104105106108109 200201202203204205206207 185187188 189191192 194 195197198 199201 5156616671768186 9095100105110115120135 161163164168171173174176 179181183185186188190192

7. 7 Introduction (c.) 103102103104 105108109 201 202 204 207 185187188 189191192 194 195197198 199201 5156616671768186 9095100105110115120135 161163164168171173174176 179181183185186188190192 Stego-Image 1010 0101 1100 0011 Secret Message

8. 8 Introduction (c.) Frequency Domain  Secret message is directly embedded into frequency domain 100101102103104105106107 200201202203204205206207 185186187188189190191192 193194195196197198199200 5055606570758085 9095100105110115120135 160162164168170172174176 178180182184186188190192 Cover Image 9772 + 0i-68 + 184i-72 + 84i-80 + 36i-84 + 0i-80 - 36i-72 - 84i-68 - 184i 759 - 763i-33 - 81i-3 - 50i18 - 39i32 - 27i41 - 9i52 + 13i72 + 48i -1486+ 54i27 + 30i10 + 22i1 + 25i-10 + 22i-23 + 14i-34 + 2i-45 - 21i -183 - 439i-14 - 15i-8 - 11i-7 - 13i0 - 15i12 - 16i23 - 10i39 + 11i -1328 + 0i-10 - 4i0 - 8i6 - 8i8 + 0i6 + 8i0 + 8i-10 + 4i -183 + 439i39 - 11i23 + 10i12 + 16i0 + 15i-7 + 13i-8 + 11i-14 + 15i -1486 -554i-45 + 21i-34 - 2i-23 - 14i-10 - 22i1 - 25i10 - 22i27 - 30i 759 + 763i72 - 48i52 - 13i41 + 9i32 + 27i18 + 39i-3 + 50i-33 + 81i Frequency Domain (Cover Image) 1010 0101 1100 0011 Secret Message 9772 + 0i-68 + 184i-72 + 84i-80 + 36i-84 + 0i-80 - 36i-72 - 84i-68 - 184i 759 - 763i-33 - 81i-3 - 50i18 - 39i32 - 27i41 - 9i52 + 13i72 + 48i -1486+ 54i27 + 30i10 + 22i1 + 25i-10 + 22i-23 + 14i-34 + 2i-45 - 21i -183 - 439i-14 - 15i-8 - 11i-7 - 13i0 - 15i13 - 16i24 - 10i39 + 11i -1328 + 0i-10 - 4i0 - 8i6 - 8i9 + 0i6 + 8i1 + 8i-10 + 4i -183 + 439i39 - 11i23 + 10i13 + 16i1 + 15i-8 + 13i-8 + 11i-14 + 15i -1486 -554i-45 + 21i-34 - 2i-24 - 14i-10 - 22i1 - 25i11 - 22i27 - 30i 759 + 763i72 - 48i52 - 13i41 + 9i32 + 27i18 + 39i-3 + 50i-33 + 81i Frequency Domain (Stego-Image) 100101102103104105106107 200201202203204205206207 185186187188189190191192 193194195196197198199200 5055606570758085 9095100105110115120135 160162164168170172174176 178180182184186188190192 Stego-Image

9. 9 Introduction (c.) Quantization Domaint  Quantization-based images such as vector quantization (VQ) Compress Codebook x x Uncompress m Codebook x x m 1 2 2 1

10. 10 Related Work  Jo and Kim’s watermarking partitions the codebook into three sub-codebooks to hide watermark  Chang and Wu’s scheme clusters codewords with difference size and performs cycle permutation  Chiang and Tsai’s scheme divides the codebook into several sub-codebooks with codeword overlapping

11. 11 Related Work (c.) Jo and Kim’s Watermarking  Jo and Kim’s technique partitions the codebook into three sub-codebooks

12. 12 Related Work (c.) Embed Cover Image 01 10 Watermark Embedded Image 01

13. 13 Related Work (c.) Chiang and Tsai´s Scheme

14. 14 Related Work (c.) Four-codeword Three-codeword Two-codeword One-codeword + Capacity(bit) 2 3 1 0 Sub-cluster

15. 15 Related Work (c.) Order in A Order in B Value in binary 00000 (0) 01001 (1) 02010 (2) 10011 (3) 11100 (4) 12101 (5) 20110 (6) 21111 (7) 22Unuse  Two codewords belongs to 3-member sub-cluster are considered together to embed three secret bits.

16. 16 Proposed Scheme  Color Image Quantization Palette Index Table Palette Color Mapping Original Quantized

17. 17 Proposed Scheme (c.)  Palette Partition Four-color Three-color Two-color One-color Capacity(bit) 2 1or 2 1 0 Sub-palette

18. 18 Proposed Scheme (c.)  Palette Partition with Overlapping Location0123 Four- color 32153916 32533947 32153147 Three- color 146137145 146147145 Two- color 125149 131149

19. 19 Proposed Scheme (c.) Location 0123 Four- color (00) 2 (01) 2 (10) 2 (11) 2 Three- color (00) 2 (1) 2 (01) 2 Two-color (0) 2 (1) 2  Lookup table for embed data

20. 20 Proposed Scheme (c.) Location0123 Four-color 32153916 32533947 32153147 Three- color 146137145 146147145 Two-color 125149 131149 Location0123 Four-color(00) 2 (01) 2 (10) 2 (11) 2 Three- color (00) 2 (1) 2 (01) 2 NA Two-color(0) 2 (1) 2 NA Quantized Color1466739149 Sub-palette Size3142 Embedded Color 147(1) 2 137(1) 2 146(00) 2 145(01) 2 67 32(00) 2 53(01) 2 15(01) 2 39(10) 2 47(11) 2 16(11) 2 125(0) 2 131(0) 2 149(1) 2 Case Quantized Color = 146 Secret Message 1110…. Sub-palette Size = 3 147 replace 146 Secret Message “1” Embedded

21. 21 Proposed Scheme (c.) Extraction Procedure Location0123 Four-color 32153916 32533947 32153147 Three- color 146137145 146147145 Two-color 125149 131149 Location0123 Four-color(00) 2 (01) 2 (10) 2 (11) 2 Three- color (00) 2 (1) 2 (01) 2 NA Two-color(0) 2 (1) 2 NA 147 Index = 147 Secret Message = 1….

22. 22 Experimental Results  RGB color images, “Airplane,” “Lena,” and “Pepper” of 512 × 512 pixels AirplaneLenaPepper Images Sub-Palette size Four-colorThree-colorTwo-color Airplane113(214)3(9)2(4) Lena146(236)2(4)1(2) Peppers71(141)5(12)22(42)

23. 23 Experimental Results (c.) Sub-palette Four- color Three- color Two- color Estimated Capacity Airplane255,2162,487893515,478 Lena251,9383,056553509,532 Peppers197,04514,94726,457445,508 Performance of the proposed method MethodCQ Jo & Kim’sChiang& Tsai’sProposed PSNR Capacity RatioPSNRCapacityRatio PSNR CapacityRatio Airplane 38.6635.58 207,966 67,52132.13512,568 78,494 32.11515,03778,632 Lena36.7233.52 206,248 64,45230.08505,957 76,198 30.11509,03677,010 Pepper33.0730.67 150,756 62,81527.28428,020 73,924 27.31442,98976,908 Average 36.1533.26 188,323 65,16429.83482,181 76,294 29.84489,02077,499 Comparison: Ratio= Capacity/(CQ-PSNR)

24. 24 Conclusions  An overlapping color palette partition based data hiding with improved data embedding procedure has been presented  Proposed method provides a largest hiding capacity and supports better stego-image quality than Chiang & Tsai’s method

26. 26 Proposed Scheme Palette Partition Algorithm  Squared Euclidean Distance (SED) Euclidean Distance Image quantization Index1 Index2

27. 27 Related Work (c.)  Chang & Wu’s Scheme  Clusters codewords with difference size  Performs cycle permutation  Embeded size in a group I = Index321539165347 Embeded data 000110110001

28. 28 Experimental Results (c.) Sub-palette Four- color Three- color Two- color Estimated Capacity Airplane255,2162,487893515,478 Lena251,9383,056553509,532 Peppers197,04514,94726,457445,508 Performance of the proposed method MethodCQ Jo & Kim’sChang & Wu’sChiang& Tsai’sProposed PSNRCapacityPSNRCapacityPSNRCapacityPSNRCapacity Airplane38.6635.58207,96632.38451,62932.13512,56832.11515,037 Lena36.7233.52206,24830.55421,20630.08505,95730.11509,036 Peppers33.0730.67150,75627.11410,58227.28428,02027.31442,989 Average36.1533.26188,32330.01427,80429.83482,18129.84489,020 Ratio65,164696,74576,29477,499 Comparison: Ratio=Capacity/(CQ-PSNR)