Proposed process of motion-activity recovery

Chapter 4 Protection of Privacy-sensitive Motion Activities in

4.2 Proposed Method for Protecting Privacy-sensitive Motion Activities in

4.2.3 Proposed process of motion-activity recovery

In this section, we describe how we retrieve the original privacy-sensitive image sequence from the protected image sequence. Before recovering each privacy-sensitive image frame in this process, we have to extract the recovery information from the corresponding protected image frame such as the start and end positions of the protected region, and regain the background image part as described in Section 4.2.1. With the background image part and the protected camouflage image, we can retrieve the original privacy-sensitive image content by the recovery process described in Algorithm 3.7. The above-described steps are illustrated in Figure 4.4.

The details are described as an algorithm in the following.

69 Source value p

Extract the recover information and recover the camouflage image

Background image

Camouflage image part

Surveillance image

Protected image

Background image part B

Stego-value q

Recover private region

Recovery sequence

Secret key

1101000……

………10110

Surveillance image

Figure 4.4 Flowchart of the private motion-activity recovery process.

Algorithm 4.3: process for the private motion-activity recovery.

Input: a protected color image sequence Sc' = {Sc1', Sc2', …, Scn'} and a protected depth image sequence surveillance S_d' = {Sd1', Sd2', …, Sdn'}; a pre-selected background color-image frame Bc, a pre-selected background depth-image frame Bd, both of the full size of the image frame; and the secret key K used in Algorithm 4.2.

Output: the original surveillance image sequence Sc and Sd recovered from Sc'' and S_d''.

Steps:

Step 1. For i = 1, 2, …, n, perform the following steps.

1.1 Extract the protected region R_i from the image contents of S_ci'' and S_di'', respectively, and retrieve the recovery sequence LRi and its length for recovering the original image contents S_c and S_d, by performing Steps 1 through 3 of Algorithm 3.7.

1.2 Cut out respectively the regions in the background images B_c and B_d corresponding to Ri as Bci' and Bdi' for use as background image parts for S_ci' and S_di', respectively.

1.3 Take Sci' and Sdi' as the camouflage images and Bci and Bdi as the original background image parts, and perform Steps 4 through 8 of Algorithm 3.7 with the key K to recover the original privacy-sensitive images Sci and Sdi..

Step 2. Output the resulting surveillance image sequences S_c and S_d.

4.3 Experimental Results

Some experimental results of applying the proposed method for protecting privacy-sensitive motion activities in a surveillance video are shown in Figures 4.5 through 4.13. In the surveillance video, we hope that the personal movement is not leaked. First, six representative images of the protected video yielded by the proposed method using Algorithms 4.2 and 3.6 are shown in Figures 4.5 through 4.7. Next, six representative images of the recovered video yielded by the proposed method using Algorithm 4.3 are shown in Figures 4.8 through 4.10. At last, we used Algorithm 3.1 to merge the color and depth images to display the 3D surveillance video. Some frames of the results are shown in Figures 4.11 through 4.13. These experimental

results show that the information of the privacy-sensitive motion activity can be protected and recovered automatically and successfully by the proposed method.

(a) (b)

(e) (f)

Figure 4.5 Six representative frames of a color surveillance video. (a) The background image frame. (b) The 19th frame. (c) The 20th frame. (d) The 21th frame. (e) The 22th frame. (f) The 23th frame.

(a) (b)

(e) (f)

Figure 4.6 Six representative frames of a depth surveillance video corresponding to that shown Figure 4.5. (a) The background image frame. (b) The 19th frame. (c) The 20th frame. (d) The 21th frame. (e) The 22th frame. (f) The 23th frame.

(a) (b)

(e) (f)

Figure 4.7 Six representative frames of a 3D surveillance video which is the result of combining those of Figures 4.5 and 4.6. (a) The background image frame. (b) The 19th frame. (c) The 20th frame. (d) The 21th frame. (e) The 22th frame. (f) The 23th frame.

(a) (b)

(e) (f)

Figure 4.8 Six representative frames of the privacy-protected color video yielded by the proposed method with Figures 4.6 and 4.6 as inputs. (a) The background image frame. (b) The 19th frame. (c) The 20th frame. (d) The 21th frame. (e) The 22th frame.

(f) The 23th frame.

(a) (b)

(e) (f)

Figure 4.9 Six representative frames of the privacy-protected depth video

corresponding to that of Figure 4.8. (a) The background image frame. (b) The 19th frame. (c) The 20th frame. (d) The 21th frame. (e) The 22th frame. (f) The 23th frame.

(a) (b)

(e) (f)

Figure 4.10 Six representative frames of the 3D privacy-protected video which comes from combination of Figures 4.8 and 4.9. (a) The background image frame. (b) The 19th frame. (c) The 20th frame. (d) The 21th frame. (e) The 22th frame. (f) The 23th frame.

(a) (b)

(e) (f)

Figure 4.11 Six representative frames of the recovered color video resulting from Figure 4.8. (a) The background image frame. (b) The 19th frame. (c) The 20th frame.

(d) The 21th frame. (e) The 22th frame. (f) The 23th frame.

(a) (b)

(e) (f)

Figure 4.12 Six representative frames of the recovered depth video resulting from Figure 4.9. (a) The background image frame. (b) The 19th frame. (c) The 20th frame.

(d) The 21th frame. (e) The 22th frame. (f) The 23th frame.

(a) (b)

(e) (f)

Figure 4.13 Six representative frames of the 3D recovered video combining the previously-shown color and depth images of Figures 4.11 and 4.12. (a) The

background image frame. (b) The 19th frame. (c) The 20th frame. (d) The 21th frame.

(e) The 22th frame. (f) The 23th frame.

Chapter 5 3D Steganography via KINECT Images

5.1 Introduction

In recent years, the various information of intellectual property becomes more and more important. Stealing the ideas of innovative creations or plagiarizing others’

works is also a kind of copyright’s violation. Even though such behaviors are difficult to define and prevent, image steganography is a feasible strategy to avoid such events.

Due to the popularity of the Internet, image data are shared frequently on the Internet.

Therefore, it is also desired to propose a method for 3D image steganography, by which a user can send secret data to other persons via the Internet or keep them securely in any digital storage.

Besides, the popularity of the 3D information which can be constructed by the KINECT device is growing dramatically. In order to reach the goal of hiding information in 3D image, we propose a method for 3D image steganography utilizing 3D images processing techniques.

In Sections 5.1.1 and 5.1.2, the related problem definitions and the idea of the proposed method are given. The proposed method for 3D image steganography by a LSB-modification scheme is presented in Section 5.2. Finally, some experimental results showing the method are given in Section 5.3.

5.1.1 Problem definition

The KINECT device has become more and more popular in recent years.

Therefore, the numbers of related applications are also growing in many fields.

What’s more is that we can not only obtain the color and depth information, but also the three-dimensional information around the environment through the KINECT device. For example, we can use the KINECT device to capture a color image and a depth image, and then transfer the depth image into the 3D coordinates for various digital applications. The acquisition of the 3D coordinates of objects is more easily than before. Image steganography of these kinds of information will become more important.

For the reason above, we want to embed a secret image into a 3D cover image with the secret key to protect the information in this study. However, such embedding might be discovered by people. One way to deal with this issue is to hide the secret image into the unused points in the 3D cover image, and add the recovery information into the resulting 3D camouflage image to generate the 3D image steganography.

Later, we can extract this information from the 3D camouflage image to recover the 3D secret image. The recovery information can be acquired only by an authorized user who has a secret key. That is, the user who has the secret key can remove the secret image from the 3D camouflage image and recover the embedded recovery information simultaneously. These conditions should be considered seriously when designing a method for protection of 3D image.

5.1.2 Proposed Idea

The basic idea of the proposed method for 3D steganography via KINECT images were represented in this section. At first, we combine the color image and the depth image together to form a 3D image. Then, we modify the pixel values of a given 3D cover image according to certain color and coordinate tables which are formatted as (x, y, z, r, g, b). Next, we not only embed the secret image into the cover image, but also embed the secret key to produce a camouflage image. In addition, only authorized users who have the secret key can extract the recovery information from the 3D image.

On the other hand, The original 3D image data we want to recover should be saved, and the recovery information for this purpose is created, called a recovery sequence. Then, we transform the resulting camouflage image into color and coordinate tables. Finally, the recovery sequence is embedded into the resulting camouflage image by a LSB-modification scheme. With this recovery sequence, the modified cover image part during the data embedding process as well as the hidden secret image can both be retrieved losslessly.

Detailed algorithms describing the proposed method and the related processes of 3D steganography are presented in the following sections.

5.2 Proposed Method for 3D

Steganography via KINECT Images

In this section, the details of the proposed method for embedding the secret image into the camouflage image by a LSB-modification scheme are described. The detailed process of preprocessing of the secret image before data hiding is described

in Section 5.2.1. In Section 5.2.2, the process of secret image hiding is described. And in Section 5.2.3, the process of secret image recovery is described.

5.2.1 Preprocessing of secret image before data hiding

The idea of preprocessing the secret image before data hiding is based on Ma and Tsai [34]. The proposed method aims to transform a group of 3D images into a 3D model where the 3D images are constructed from the color and depth images acquired with a so-called octagonal 9-KINECT imaging device. The device, which consists of nine KINECT devices, can scan objects or an environment from top to bottom in a raster scan order, so that it can use the difference of height values to filter out the floor from each acquired depth image.

In addition, the unchanging nature of the static environment facilitates us to find out the position of the object in the acquired KINECT images. Therefore, in constructing an object model, we can remove undesired surrounding scene parts by eliminating those pixels with larger depth values. For example, we have used a toy bear as the object. In the recording procedure, the bear was pushed forward slowly by a person who keeps a fixed distance with respect to the bear. The fixed distance enables us to segment out the bear easier. Subsequently, we superimpose the bear appearing in every two frames at a time by use of a so-called distance-weighted correlation (DWC) measure to get a complete 3D image of the bear. In order not to choose duplicate points, we integrate all the points of the bear together without repeating identical points. An in the resulting point group, we choose every three neighboring points to form a surface, resulting in a model of the bear finally.

Now, we briefly review the measure of DWC which was proposed by Fan and Tsai [35] originally for automatic Chinese seal identification. The measure is defined as the minimum distance between two groups S and T of pixels of seal imprint images after the seal imprint images are overlapped. For each pixel p in S, a pixel in T with the minimum distance to p is searched for. If the result is a pixel p' within a limited circular area A with a pre-selected radius K, then a weight wpK

= 1/(dp2 used to decide an effective distance; distances larger than this threshold are discarded.

Finally, the DWC defined for the two groups of pixels, S and T, is defined as follows:

where the coefficient 1/2 is included to treat S and T symmetrically; and NS and NT are the total numbers of pixels in S and T, respectively. It can be verified that 0  C^K  1 and C^K = 1 if and only if S = T. The DWC, though defined originally for seal identification, is a general measure for point-type object shape matching.

In the method proposed in this study, an extension of the above-reviewed 2D DWC measure, called 3D DWC, is used for the purpose of 3D model construction used in constructing the 3D secret message which is an object. The 2D DWC was used to decide whether the images of two object shapes are similar or not, while the 3D DWC is used to compute the displacement of two similar objects.

5.2.2 Review of the reversible contrast mapping (RCM) for lossless data hiding

We use the RCM technique proposed by Coltuc and Chassery [32] for data hiding in the proposed 3D steganography, which we review here.

For a pair of pixel values, (x1, x2), three cases should be handled: (a) if 0  2x1 

x₂  255 and 0  2x2  x₁  255, and if either x1 and x₂ is not an odd value, then we transform (x₁, x₂) into (x₁', x₂') by x₁' = 2x₁x2 and x₂' = 2x₂x1, set the LSB of x₁' to be “1,” and hide the bit in the LSB of x2'; (b) if 0  2x1 x2  255 and 0  2x2 x1  255, and if both x₁ and x₂ are odd values, then we set the LSB of x₁ to “0” and hide the bit in the LSB of x2; (c) if (x1, x2) do not satisfy both constraints of 0  2x1  x2  255 and 0  2x2  x1  255, then we set the LSB of x1 to “0” and append the original LSB of x₁ to the end of the data string for hiding in order to losslessly recover the original LSB of x1 later during the extraction process.

5.2.3 Proposed secret image hiding process

The proposed secret image embedding process for 3D image steganography by a LSB-modification scheme is described in this section. In the embedding process, we first transform a group of 3D images into a single 3D model where the 3D images are constructed from the color and depth images acquired with a so-called octagonal 9-KINECT imaging device. As mentioned before, we use the distance-weighted correlation (DWC) measure to get this complete 3D model of an object. Consequently, we format the coordinates of the 3D model as (x, y, z, r, g, b).

In addition, we also formulate the information of the secret message S in a 3D fashion as (x, y, z, r, g, b), such that we can embed the secret message S into a cover image C in a easier way. Then, we transform the data, add a recovery sequence, and

use a secret key K to randomize the order of the bits in the recovery sequence.

In more detail, during the data hiding process, we hide the secret image S into the unused points in the cover image C. Next, we mark each point with a label to record whether the points of the cover image C is changed or not. After these processes, the recovery information for this purpose is created, called a recovery sequence. Finally, the recovery sequence which records the label of each point is embedded into the camouflage image by a lossless LSB-modification scheme called RCM (Reversible contrast mapping) [32], and then a stego image is finally generated.

A flowchart of the proposed secret image embedding process is given in Figure 5.1 and an algorithm describing the process is given below as Algorithm 5.1. We assume that the 3D secret image S is an object or a group of points within a limited 3D space range so that it can be enclosed in the space of a cube with a certain size.

Also, we assume that there exist groups of non-object points forming “holes” of cubic shapes which are large enough to contain the 3D secret image.

Algorithm 5.1: secret image embedding for image steganography.

Input: a 3D secret image S and a 3D cover image C with assumptions described above; and a secret key K.

Output: a 3D stego-image V into which S is embedded.

Steps.

Stage 1 --- Search of a “hole” in the 3D cover image larger enough to contain the 3D secret image.

Step 1. Conduct the following steps to estimate the size of the 3D secret image S.

1.1 Make a cube Bs with size rs  rs  rs.

1.2 Check if all the points of the secret image S are enclosed in the cube: if not, increment r_s and repeat Step 1.1; else, continue.

1.3 Take the final cube Bs0 and its size rs0  rs0  rs0 as the enclosing cube of S.

Cover Image

Secret Image

Stego Image Camouflage

Image

Secret key Random process

Recovery sequence

Randomized Camouflage

Image Embed process

Figure 5.1 A flowchart of the proposed secret image embedding process.

Step 2. Search the 3D cover image C for a “hole” which is large enough to contain the 3D secret image S by the following steps.

2.1 Find an unvisited non-object point P_c in C in a raster scan order.

2.2 Make a cube Bc with size rc  rc  rc with Pc as the center and with the initial size set to be 333.

2.3 Check if all the points of the 3D cover image C within the cube are non-object points: if not, go to Step 2.1 to find another hole; otherwise, take the current cube Bc with size rc  rc  rc as a hole H.

2.4 Compare the size rs0  rs0  rs0 of the enclosing cube Bs0 of S with the size r_c  rc  rc of the hole H (i.e., compare r_s0 with r_c) to see if H can contain the cube B_s0 of S: if so, go to Step 3; else, enlarge the hole by increment rc by one and go to Step 2.2.

Stage 2 --- Embedding the 3D secret image into the found “hole” in the 3D cover image.

Step 3. Embed S into the current cube B_c of C with the size r_c  rc  rc with the key K to produce a camouflage image C_a as follows.

3.1 Move S to the center of the cube B_c in C, which can contain S.

3.2 Record in order the side length r_c and the coordinates (x_c, y_c) of the center point P_c of the cube B_c as a recovery sequence L_R.

3.3 Randomize the positions of the pixels in S by the secret key K to produce a camouflage image C_a.

Step 4. Perform the following steps to embed the recovery sequence LR into Ca to produce a 3D stego-image V.

4.1 Transform LR into a binary string SR, count the number N of bits in SR, and convert N into a 18-bit binary string as N_R.

4.2 Combine strings NR and SR by adding NR to the front of string SR. 4.3 Embed the first three unembedded bit b_s of S_R respectively into the R,

G, and B color values of a pair of object pixels in Ca selected a raster-scan order (with all the bits in S_R regarded as unembedded initially) by the lossless RCM scheme [32] as reviewed in Section 5.2.2.

4.4 Repeat Step 4.3 until all the bits in SR are embedded in Ca. Step 5. Output the resulting Ca as the 3D stego-image V.

Note that in Step 4.3 above, while carrying out the RCM scheme, if case (c) with the pair of pixel values being x1 and x2 is encountered as described in Section 5.2.2, the LSB of x₁ should be changed to be “0,” and the original LSB of x₁ should be appended to the end of the data string SR for hiding.

5.2.4 Proposed secret image recovery process

In the secret image recovery process, we have two input sets of data for the image retrieval work. The first is a stego-image, and the second is a user key. If the given user key is different from the one which is used in the secret image embedding process, the extraction of the secret image will fail. By using a reverse version of the RCM scheme and the right key, we can recover the original secret image from the stego-image. The detailed steps and a flowchart of the proposed secret image recovery process are given in Algorithm 5.2 and Figure 5.2, respectively.

Algorithm 5.2: secret image recovery.

Input: a 3D stego-image V and a secret key K.

Output: a recovered 3D secret image S.

Steps.

Step 1. Perform the following steps to extract the length of recovery sequence LR

from V.

1.1 Select an object pixel pair, T1 and T2, from V in a raster-scan order.

1.2 Extract the three LSBs respective from the R, G, and B color values of T1 and T2 by an inverse version of the lossless RCM scheme [32] and append them into a string N_R (initially empty).

1.3 If the length of sequence NR is not equal to 18, then go to Step 1.1 to extract more bits to compose N_R; otherwise, continue.

91 Secret Image

Stego Image

Camouflage Image

Secret key Extract recovery

sequence

Recover process Recovery

information

Figure 5.2 Flowchart of proposed secret image recovery process.

Step 2. Transform N_R into a decimal integer L and regard it to be the length of the recovery sequence LR to be extracted next.

Step 3. Perform the following steps to extract the recovery sequence L_R from V.

3.1 Select an unprocessed object pixel pair T1 and T2 from V in a raster-scan order.

3.2 Extract three LSBs respectively from the R, G, and B color values of T1

and T₂ by an inverse version of the lossless RCM scheme [32] and append them to a string SR (initially empty).

3.3 If the length of string S_R is not equal to L, then go to Step 3.1;

otherwise, continue.

Step 4. Transform the extracted string SR into the recovery information of a side length r_c and the coordinates (x_c, y_c) of the center point P_c of a cube B_c within the stego-image V.

Step 5. According to the side length r_c and the coordinates (x_c, y_c), find the cube B_c with the size r_c  rc  rc and the center point at coordinates (x_c, y_c) in V, and

在文檔中利用3D KINECT影像做視訊監控應用上之隱私保護與秘密隱藏 (頁 80-0)