Chapter 2 Review of Related Works
2.3 Previous Studies on Information Hiding Techniques via Art Images
The combination of information hiding techniques and art image creation is a new idea of information security technology. Techniques based on this idea utilize the characteristics of the creation process of the art image to embed extra information in the generated images. Due to this way of camouflage, secret data can so be kept or transmitted covertly and securely. In addition, hackers will also tend to get unaware of the secret embedded in such images and this reduces the danger of being stolen or being tampered with.
Specifically, Lin and Tsai [15] proposed algorithms to embed secret messages in image mosaics by adding visible boundary regions to the four sides of tiles and modifying the histogram of tile images. Wang and Tsai [16] presented a data hiding technique for image mosaics as well. By varying the overlapping degrees of adjacent tile images, the method can create a new-style mosaic image in which bits of the message data are embedded. Some resulting images created via these two methods are shown in Figure 2.6. Different from the intuitive idea of image mosaics, Lai and Tsai [17] created a new type of mosaic image, called secret-fragment-visible mosaic, which is reconstituted with rectangle fragments yielded by partitioning of the original image. A method to embed secret messages is proposed by switching the relative positions of tile images which have similar colors in an identical bin of the histogram.
The resulting mosaic image is still a meaningful image like another one, as shown in
19
Figure 2.7.
In addition to previous methods, numerous researches on combining other types of art images and data hiding have been given. Hsu and Tsai [18] presented three new types of art images and three methods to hide secret information in art images by using the features of the creation process. The first type of image, digital puzzle image, is generated to embed data by modifying the orientations, sizes, and angles of the
(a) (b)
Figure 2.6 Image mosaics. (a) An image mosaic created from Lin and Tsai [15]. (b) An image mosaic created from Wang and Tsai [16].
(a) (b)
Figure 2.7 A Secret-fragment-visible mosaic image created with Lai and Tsai’ method [17]. (a) Original image. (b) Generated secret-fragment-visible mosaic image.
20
puzzle pieces. Second, in the new type of pointillistic image, palette colors are used for data hiding by varying the RGB values of each color dot of the pointillistic image.
And the last, a new art image called circular-dotted image is created to embed secret messages by using the drawing order of the circular dots and a circular dot overlapping scheme. Some examples of the art images created by Hsu and Tsai [18]
are shown in Figure 2.8.
Additionally, an information hiding approach was proposed through the use of stained glass images by Hung and Tsai [19]. The secret data can be hidden in stained glass images by modifying the tree structure used in the creation process. A result generated by the method is shown in Figure 2.9(a). Chang and Tsai [20] created a new type of art image, called tetromino-based mosaic, which is composed of tetrominoes of the Tetris game. A tetromino is a geometric shape composed of four squares which is connected orthogonally. By the composition of geometric forms, tetrominoes can be combined to fit into a fixed shape (rectangles mostly) to form blocks which then are used to fill an image plane. A data hiding method is proposed by using distinct
(a) (b) (c)
Figure 2.8 Art images created by Hsu and Tsai [18]. (a) A digital puzzle image. (b) A digital pointillistic image. (c) A digital circular-dotted image.
21
combinations and color shifting of the tetromino elements. An image yielded by Chang and Tsai [20] is shown in Figure 2.9(b).
In this study, we also propose new methods which combine information hiding techniques and art image creation to achieve covert communication. By utilizing the characteristics of the creation processes of three art images, which are line-based Cubism-like image, strip-based Futurism-like image, and rectangle-based Neo-Plasticism-like image, the images can be transmitted or kept with the secret data embedded without arousing attention from other people.
(a) (b)
Figure 2.9 Two examples of art images. (a) A stained glass image from Hung and Tsai [19]. (b) A tetromino-based mosaic from Chang and Tsai [20].
22
Chapter 3
Line-based Cubism-like Image --- A New Type of Image and Its
Application to Data Hiding by Invisible Reversible Pixel
Re-coloring
3.1 Overview of Proposed Method
In this chapter, we describe how we create a type of art image like Cubism paintings automatically via the use of a computer, and we name this kind of art image
line-based Cubism-like image. By this type of art image, we try to keep a
characteristic of the Cubism art multiple viewpoints by the use of the line
feature. By rearranging lines in a given image, which are yielded by applying the Hough transform to the image, a line-based Cubism-like image is created, which includes a new three-dimensional shape of each identity in the given image. In Section 3.2, the proposed method for automatic creation of line-based Cubism-like images will be described in detail.In order to achieve the purpose of hiding information in this type of art image, we propose also a data hiding technique in this study. A given message is embedded into a line-based Cubism-like image during the stage of region coloring in the creation process of the image. We assign a new color to each image pixel by keeping unchanged the average of the color in the region which includes the pixel, and
23
re-coloring the pixel without causing a perceptible change. Furthermore, a technique is proposed to enhance the security of the hidden data by randomizing the processing order of the regions.
3.2 Proposed Line-based Cubism-like Image Creation Process
3.2.1 Idea of Proposed Creation Technique
Cubism artists transform a natural scene into geometric forms by breaking up, analyzing, and re-assembling objects in the scene. In addition, with the scene objects rearranged to intersect at random angles, each painting of Cubism seems to be composed of intersecting lines and fragmented shapes in an abstract style. The idea of the proposed art image creation method is inspired by this concept of Cubism, as mentioned previously.
In the creation process of a line-based Cubism-like image from a given image, at first we find the longer line segments in the source image by the Hough transform.
Then, we connect the line segments and extend them to reach the image boundaries.
Finally, we generate the desired art image via the operations of line segment merging and region re-coloring. This process accomplishes the goal of transforming the input image into an abstract form since the lines of the created Cubism-like image tend to constitute the skeleton of the objects in the source image as observed from according to our experimental results. The detailed algorithms of the above-mentioned processes are described in the following sections.
24
3.2.2 Proposed Art Image Creation Process
In this section, we present an algorithm which implements the idea of proposed Cubism-like image creation. Basically, in the process of line detection, we find edges of the source image by utilizing the Canny edge detection technique [21], and then perform the Hough transform on the edge detection result to obtain longer line segments in the source image. By extending and recombining these longer line segments, a desired Cubism-like image is created. The detailed algorithm is given as follows.
Algorithm 3.1: line-based Cubism-like image creation.
Input: a source image S, and two threshold values the minimum length L
min of a line segment, and the minimum distance Dmin between two lines.Output: a line-based Cubism-like image C.
Steps.
Stage 1 --- creating crossing-image lines.
Step 1. Perform Canny edge detection to find the edges E1, E2, …, En in source image S, resulting in a new image S′.
Step 2. Implement the following steps to find out longer line segments in S′.
2.1 Find the line segments L1, L2, …, Lm, in S′ by applying the Hough transform on S′, yielding a second new image S′′ of the line type.
2.2 Select those line segments in S′′ with their lengths larger than the threshold Lmin.
2.3 Compare every line pair Li and Lj with i j in S′′ in the following way:
if the distance Dij between Li and Lj is smaller than Dmin, then delete Li if the length of Li is smaller than that of Lj; or delete Lj, otherwise.
25
Step 3. Extend each of the remaining line segments in S′′ to the boundaries of S′′, and regard the source image S as being partitioned by the extended lines into regions.
Stage 2 --- re-coloring image regions.
Step 4. Create a binary image T with the same size as that of S with the initial pixel values all set to be 0.
Step 5. Fill the value of 1 into those pixels in T which correspond to those lying on each of the extended line segments in S′′.
Step 6. Implement following steps to recolor the regions in S.
6.1 Perform region growing in the binary image T in a raster-scan order, and segment out 0-valued regions, R1, R2, …, Rk, each of which is enclosed by a group of 1-valued line segments in S′′.
6.2 Compute the area Ai of each segmented region Ri in T and the average RGB color values (Cir, Cig, Cib) of the corresponding region Ri′ in S using Ai, and re-color each pixel in Ri′ of S by the color values (Cir, Cig,
C
ib), i = 1, 2, …, k.6.3 Re-color all lines in S corresponding to the 1-valued extended lines in
T by the white color.
Step 7. Take the final S as the desired line-based Cubism-like image C.
The above algorithm of line-based Cubism-like image creation, as illustrated in Figure 3.1, is composed of two stages. In Stage 1, we perform line detection to obtain the longer lines in a source image S. By Canny edge detection, we get a group of edge points. For the purpose of finding prominent line features in S, we use two thresholds to select the longer and sufficiently-separate lines from those line segments yielded by applying the Hough transform to the group of edge points. The first threshold Lmin is
26
used to filter out short line segments. The other threshold Dmin is used to filter out extended lines which are too close to other longer lines. The final step in this stage is to extend each of the remaining line segments to cross the image, with the two line
Figure 3.1 Process of crossing-image line creation.
In this study, after considering the mutual influence between the image size and the line length, we use one-tenth of the longer boundary length of the image as the initial value of Lmin and Dmin. A series of experiments about the effects of varying the values of Lmin and Dmin have been conducted, and an experimental result is shown in Figure 3.2. In these resulting images, we can find that if we take a smaller initial value
27
of Lmin, the number of extracted lines will increase. With more lines, the complexity of the created Cubism-like image also increases, giving an impression closer to the original image content. On the other hand, fewer lines make the Cubism-like image simpler and more abstract. The effect of changing the initial value of Dmin is similar to that of Lmin.
(a) (b)
(c) (d)
(e) (f)
Figure 3.2 An experimental result of varying the threshold values of Dmin and Lmin. (a) A source image with size 1024768. (b) A Cubism-like image created from (a) with initial Dmin = 102 and initial Lmin = 102. (c) A Cubism-like image created from (a) with Dmin = 20 and Lmin = 102. (d) A Cubism-like image created from (a) with Dmin = 102 and Lmin = 20. (e) A Cubism-like image created from (a) with Dmin = 200 and Lmin = 102. (f) A Cubism-like image created from (a) with Dmin = 102 and Lmin = 200.
28
In Stage 2, with the extended line segments, the source image S is regarded as being partitioned into regions. By region growing, we segment out these regions and calculate the area and the average RGB color of each of them. Finally, a line-based Cubism-like image C is created by re-coloring these regions with the average color and all the lines with the white color.
3.2.3 Experimental Results
According to the above discussions, we see that different selections of the two threshold values Lmin and Dmin will result in totally different effects. However, it is difficult to decide which result is better than the others because the decision is obviously dependent on the different feelings of people for art. Therefore, in this study we just offer a series of results yielded by the use of different sets of thresholds for the user to choose. Specifically, we use the normalized initial thresholds of 1/10 of the length of the longer image boundary as the center, and vary each threshold to be twice and half of its initial value, in addition to the initial one. As a result, each threshold has three choices, resulting in nine choices of the two threshold values.
Then, we generate nine art images, each corresponding to one of the nine threshold combinations, for the user to choose his/her favorite one among them.
Besides, we also provide the option of choosing normalized thresholds for users, and then we produce the nine sets of threshold combinations as described above based on the choice of the user. Some Cubism-like images created by the above-proposed algorithms with nine results yielded by the use of different threshold combinations for each input source image are given in Figures 3.3 through 3.5. For simplification, we use the expression (Dmin, Lmin) to show a combination of the two thresholds in the captions of the figures. As seen in the images, we can see an abstract style of Cubism
29
shown by the created lines and regions.
(a) (b)
(c) (d) (e)
(f) (g) (h)
(i) (j) (k)
Figure 3.3 Experimental results. (a) A source image with size 1024768. (b) Initial Dmin = 102 and initial Lmin = 102. (c) (Dmin, Lmin) = (51, 51). (d) (Dmin, Lmin) = (51, 102). (e) (Dmin, Lmin) = (51, 204). (f) (Dmin, Lmin) = (102, 51). (g) (Dmin, Lmin) = (102, 102). (h) (Dmin, Lmin) = (102, 204). (i) (Dmin, Lmin) = (204, 51). (j) (Dmin, Lmin) = (204, 102). (k) (Dmin, Lmin) = (204, 204). (l) A better choice of 9 images to fit the abstract style of Figure 3.3(a) is Dmin=102 and Lmin=51.
30
(l)
Figure 3.3 Experimental results. (a) A source image with size 1024768. (b) Initial Dmin = 102 and initial Lmin = 102. (c) (Dmin, Lmin) = (51, 51). (d) (Dmin, Lmin) = (51, 102). (e) (Dmin, Lmin) = (51, 204). (f) (Dmin, Lmin) = (102, 51). (g) (Dmin, Lmin) = (102, 102). (h) (Dmin, Lmin) = (102, 204). (i) (Dmin, Lmin) = (204, 51). (j) (Dmin, Lmin) = (204, 102). (k) (Dmin, Lmin) = (204, 204). (l) A better choice of 9 images to fit the abstract style of Figure 3.3(a) is Dmin=102 and Lmin=51. (Continued.)
(a) (b)
(c) (d) (e)
Figure 3.4 Experimental results. (a) A source image with size 1024768. (b) Initial Dmin = 102 and initial Lmin = 102. (c) (Dmin, Lmin) = (51, 51). (d) (Dmin, Lmin) = (51, 102). (e) (Dmin, Lmin) = (51, 204). (f) (Dmin, Lmin) = (102, 51). (g) (Dmin, Lmin) = (102, 102). (h) (Dmin, Lmin) = (102, 204). (i) (Dmin, Lmin) = (204, 51). (j) (Dmin, Lmin) = (204, 102). (k) (Dmin, Lmin) = (204, 204). (l) A better choice of 9 images to fit the abstract style of Figure 3.4(a) is Dmin=102 and Lmin=51.
31
(f) (g) (h)
(i) (j) (k)
(l)
Figure 3.4 Experimental results. (a) A source image with size 1024768. (b) Initial Dmin = 102 and initial Lmin = 102. (c) (Dmin, Lmin) = (51, 51). (d) (Dmin, Lmin) = (51, 102). (e) (Dmin, Lmin) = (51, 204). (f) (Dmin, Lmin) = (102, 51). (g) (Dmin, Lmin) = (102, 102). (h) (Dmin, Lmin) = (102, 204). (i) (Dmin, Lmin) = (204, 51). (j) (Dmin, Lmin) = (204, 102). (k) (Dmin, Lmin) = (204, 204). (l) A better choice of 9 images to fit the abstract style of Figure 3.4(a) is Dmin=102 and Lmin=51. (Continued.)
In Figures 3.3 through 3.5, Figure 3.3(a), 3.4(a), and 3.5(a) are the source images, and Figures 3.3(b), 3.4(b), and 3.5(b) show the results generated with initial thresholds. Figures 3.3(c) through (k), 3.4(c) through (k), and 3.5(c) through (k) are the experimental results with nine combinations of the thresholds. Finally, Figures 3.3(l), 3.4(l), and 3.5(l) are better choices from the respective nine results.
32
(a) (b)
(c) (d) (e)
(f) (g) (h)
Figure 3.5 Experimental results. (a) A source image with size 7681024. (b) Initial Dmin = 102 and initial Lmin = 102. (c) (Dmin, Lmin) = (51, 51). (d) (Dmin, Lmin) = (51, 102). (e) (Dmin, Lmin) = (51, 204). (f) (Dmin, Lmin) = (102, 51). (g) (Dmin, Lmin) = (102, 102). (h) (Dmin, Lmin) = (102, 204). (i) (Dmin, Lmin) = (204, 51). (j) (Dmin, Lmin) = (204, 102). (k) (Dmin, Lmin) = (204, 204). (l) A better choice of 9 images to fit the abstract style of Figure 3.5(a) is Dmin=102 and Lmin=102.
33
(i) (j) (k)
(l)
Figure 3.5 Experimental results. (a) A source image with size 7681024. (b) Initial Dmin = 102 and initial Lmin = 102. (c) (Dmin, Lmin) = (51, 51). (d) (Dmin, Lmin) = (51, 102). (e) (Dmin, Lmin) = (51, 204). (f) (Dmin, Lmin) = (102, 51). (g) (Dmin, Lmin) = (102, 102). (h) (Dmin, Lmin) = (102, 204). (i) (Dmin, Lmin) = (204, 51). (j) (Dmin, Lmin) = (204, 102). (k) (Dmin, Lmin) = (204, 204). (l) A better choice of 9 images to fit the abstract style of Figure 3.5(a) is Dmin=102 and Lmin=102. (Continued.)
3.3 Proposed Technique for Data Hiding in Line-based Cubism-like Images by Invisible Reversible Pixel
Re-coloring
3.3.1 Idea of Proposed Data Hiding Technique
The proposed method of using Cubism-like images for covert communication is
34
described in this section. In the proposed Cubism-like image creation process as presented by Algorithm 3.1 above, we re-color image regions with the respective average colors of the regions. Due to the nature of the human visual system, people cannot sense small changes in the appearance of a color image, such as color alternations or edge shiftings. Accordingly, we implement a method to hide secret a message in a cover image (a cubism-like color stego-image generated by the proposed method described previously) by slightly changing the RGB color values of the pixels in each region of the cover image. As a result, people will not be able to distinguish between the cover image and a stego-one. It is in this way that we achieve the goal of data hiding in the proposed line-based Cubism-like art image.
Besides, for the reason of reversibility in the hidden data extraction process, a region re-coloring technique is proposed, which keeps the average color of each region unchanged. Consequently, we can restore the color information of the pixels of the stego-image, and extract the secret messages embedded in them. In the following sections, these mentioned techniques used in the proposed method for data hiding in the line-based Cubism-like image will be described in detail.
More specifically, in the proposed data hiding process, after the step of hiding the message bits into a color channel, the pixel colors in a region will be changed via color shifting, and the average color of the region will also be influenced. In order to keep the average color unchanged, we must limit the number of embedded message bits. For this purpose, it is found in the study that the property of rounding-off may be utilized. Specifically, when computing the average color C of a region, all the computed results in the range between C
0.5 and C + 0.5 will be rounded to be an
identical value since RGB color values used in this study are integer numbers.Accordingly, we can acquire the maximum number of embedded message bits in a region R by an equation derived as follows.
35
Furthermore, we hide the secret message bits by shifting the average R-color value of each pixel in this study. We assume that when hiding a bit of 0 into a pixel P in region
R with average R-color value C
r, the pixel’s color Cr is decreased by 1; and when values of the two numbers N1 and N0 according to the above-mentioned rounding-off property, resulting in the following formula:36 pixels in the region, implying that the data hiding capacity is also restricted by it. In the best case, we know that the maximum number of embeddable bits in a region is just the number A of pixels in the region, that is,
1 0 get the maximum number of embeddable bits in a region. In the extreme case where the digit sequence is composed of all 0’s or all 1’s, the upper bound of the data hiding capacity will be reached, which is
2
A
1, as can be figured out from Equation
(3.1).
In this study, we keep the average region color unchanged for two reasons. The first, as mentioned previously, is to make possible recovery of the region information in the data extraction process, where we use the average color as a basis to extract the hidden secret message. The detail can be seen in Section 3.3.3. The other reason is to yield a visual mosaic effect such that the region color that people see is almost the
same, despite of the color shifting inside the region.
37
3.3.2 Proposed Data Hiding Process
Based on the creation process of the line-based Cubism-like image, we propose a data hiding method in this section. According to the last step of Algorithm 3.1 described previously, we re-color each region of the input image by the average color of the region to generate a Cubism-like image. For the purpose of hiding data in a generated Cubism-like image, we try to modify this re-coloring process to hide secret messages. Specifically, the data hiding process is composed of two main stages. First at all, we transform the secret massage into a digit sequence and append an ending pattern (with at least one digit) at the end of the digit sequence to keep the sequence length a multiple of three, as shown in Figure 3.6. By the ending pattern, we can determine where the massage ends in a sequence of extracted bits in the message extraction process. Then, we obtain the information of two parameters of each region, namely, the region area and the average RGB color values in the region, by performing Algorithm 3.1.
Furthermore, we use a secret key to randomize the order of re-coloring of the regions in the input image, and take the resulting new sequence as the order for data hiding. For each region, we compute the maximum data hiding capacity first. In order to keep the average RGB color of the region unchanged, we limit the embedded amount of message bit data in each region. After getting the maximum data hiding capacity, we embed the message bits by shifting the RGB color values of the pixels in each region according to the above-mentioned data hiding order. After the digit sequence is exhausted, there might exist regions in which no message bit is embedded.
We have to deal further with these intact regions to keep the coloring style of all
We have to deal further with these intact regions to keep the coloring style of all