Multimed Tools Appl– Based <strong>on</strong> the characteristics of HVS and our extensive experiments, embeddingsecret data into the edge regi<strong>on</strong>s can not also produce <str<strong>on</strong>g>more</str<strong>on</strong>g> visual indistinguishableresults but also enhance the robustness against the statistical analysis.Therefore we should make full use of those edge regi<strong>on</strong>s as far as possible.Typical PVD-<str<strong>on</strong>g>based</str<strong>on</strong>g> approaches can embed <str<strong>on</strong>g>more</str<strong>on</strong>g> secret bits into busy areas.However, lots of smooth regi<strong>on</strong>s will also be c<strong>on</strong>taminated after data embeddingeven the difference of the two c<strong>on</strong>secutive <strong>pixel</strong> <strong>value</strong>s is zero, for instance, 6bits are embedded using the method [23], while many available edges regi<strong>on</strong>shave not been fully exploited. To overcome the limitati<strong>on</strong>, we employ anopti<strong>on</strong>al threshold T to measure the edge strength for each embedding unit.When performing data embedding, our method can first estimate whether theembedding units whose strength larger than T are enough for a given secretmessage. If not, the method decreases the threshold T to T − 1 to release <str<strong>on</strong>g>more</str<strong>on</strong>g>spaces, until all the secret message can be embedded completely. In such a way,most smooth/falt regi<strong>on</strong>s in cover images can be well preserved. In this paper, Tdecreases from 32 to 1. Therefore, <strong>on</strong>ly 5 (2 5 = 32) bits additi<strong>on</strong>al informati<strong>on</strong>is needed. Please note that the threshold T is dependent <strong>on</strong> image c<strong>on</strong>tentsand the secret message to be embedded. Thus we need the threshold T asside informati<strong>on</strong> for subsequent data extracti<strong>on</strong>. In practice, we can embed Tinto a preset regi<strong>on</strong> that does not be used for data hiding using some existingsteganographic approaches, such as the simple LSB replacement.3 Proposed method3.1 Data embedding– Step 1 The cover image is first divided into n<strong>on</strong>-overlapping squares with sizeBz × Bz,whereBz is a multiple of 3, in this paper, Bz = 6. For each square, werotate it by a random degree of 0, 90, 180 and 270 <str<strong>on</strong>g>based</str<strong>on</strong>g> <strong>on</strong> a secret key key 1 ,which produces a random degree sequence with size of [ M Bz ][ N Bz ],whereM × Nis the size of the cover image.– Step 2 The resulting image is rearranged as a row vector running through all therows in a raster scanning manner as it did in the paper [22]. And then the vector ispartiti<strong>on</strong>ed into n<strong>on</strong>-overlapping embedding units with three c<strong>on</strong>secutive <strong>pixel</strong>s,say [g i , g i+1 , g i+2 ].– Step 3 For a given embedding unit, calculate two differences of the adjacent <strong>pixel</strong><strong>value</strong>s byd 1 = g i+1 − g i , d 2 = g i+2 − g i+1If |d 1 |≤T and |d 2 |≤T, 1 then skip to the next embedding unit in a pseudorandomtravel order <str<strong>on</strong>g>based</str<strong>on</strong>g> <strong>on</strong> a secret key key 2 .Otherwise, if |d 1 | > T or |d 2 | > T, which means that the middle <strong>pixel</strong> g i+1 is1 T is a threshold, the initialize setting of T is 32 in this paper.
Multimed Tools Appllocated at the c<strong>on</strong>tent edges whose strength degree is larger than T, theng i+1can be used for data embedding:Determine the range of centre <strong>pixel</strong> in the embedding unit, denoted as range g ′i+1,which satisfies that the sort order of g i , gi+1 ′ , g i+2 in the stego is the same asthe sort order of g i , g i+1 , g i+2 in the cover, also the relati<strong>on</strong>ships between d 1 , d 2(d ′ 1 , d′ 2 )andT are preserved after data hiding. For instance, if g i ≥ g i+1 ≥ g i+2in the cover, then we have g i ≥ gi+1 ′ ≥ g i+2 in the stego, if |d 1 |=|g i+1 − g i | > T,then we keep |d ′ 1 |=|g′ i+1 − g i| > T.The range can be obtained according to the following four cases:– Case 1 g i+1 < g i and g i+1 < g i+2 ,case 1.1:if |d 1 | > T & |d 2 | > Trange g ′i+1= [ 0,...,min ( g i − T − 1, g i+2 − T − 1 )]case 1.2:case 1.3:if |d 1 | > T & |d 2 |≤Trange g ′i+1= [ max ( g i+2 − T, 0 ) ,...,min ( g i − T − 1, g i+2 − 1 )]if |d 1 |≤T & |d 2 | > Trange g ′i+1= [ max ( g i − T, 0 ) ,...,min ( g i+2 − T − 1, g i − 1 )]– Case 2 g i+1 > g i and g i+1 > g i+2 ,case 2.1:if |d 1 | > T & |d 2 | > Trange g ′i+1= [ max ( g i + T + 1, g i+2 + T + 1 ) ,...,255 ]case 2.2: if |d 1 | > T & |d 2 |≤Trange g ′i+1= [ max ( g i + T + 1, g i+2 + 1 ) ,...,min ( g i+2 + T, 255 )]case 2.3:if |d 1 |≤T & |d 2 | > Trange g ′i+1= [ max ( g i+2 + T + 1, g i + 1 ) ,...,min ( g i + T, 255 )]– Case 3 g i ≥ g i+1 ≥ g i+2case 3.1:case 3.2:case 3.3:if |d 1 | > T & |d 2 | > Trange g ′i+1= [ g i+2 + T + 1,...,g i − T − 1 ]if |d 1 | > T & |d 2 |≤Trange g ′i+1= [ g i+2 ,...,min ( g i+2 + T, g i − T − 1 )]if |d 1 |≤T & |d 2 | > Trange g ′i+1= [ max ( g i+2 + T + 1, g i − T ) ],...,g i– Case 4 g i ≤ g i+1 ≤ g i+2case 4.1:if |d 1 | > T & |d 2 | > Trange g ′i+1= [ g i + T + 1,...,g i+2 − T − 1 ]