CONSTRUCTION OF LOW-DISTORTED MESSAGE-RICH VIDEOS FOR PERVASIVE COMMUNICATION

Transcription

1 2016 International Computer Symposium CONSTRUCTION OF LOW-DISTORTED MESSAGE-RICH VIDEOS FOR PERVASIVE COMMUNICATION 1 Zhen-Yu You ( ), 2 Yu-Shiuan Tsai ( ) and 3 Wen-Hsiang Tsai ( ) 1 Institute of Information Science and Engineering 2 Computer Vision Research Center 3 Department of Computer Science National Chiao Tung University, Hsinchu, Taiwan s: ystsai@g2.nctu.edu.tw, whtsai@cis.nctu.edu.tw ABSTRACT Many types of multimedia can be used to accommodate information for pervasive communication. Messages may be injected into them by information hiding techniques and extracted for various purposes. Such multimedia with messages embedded are called message-rich multimedia. A new method for creating message-rich s is proposed in this study for people to exchange messages by display-and-imaging operations. A message is firstly embedded in a given by a technique of injecting peak values into the DFT spectrum of each frame. The resulting message-rich, when displayed on a TV or monitor, is then recorded with a smart-phone camera as another. The message in the recorded is finally extracted for visual inspection. This way of message transmission can be used for covert communication, secret message exchanges, etc. Good experimental results show the feasibility of the proposed method. Keywords: low-distorted, message-rich, pervasive communication, multimedia, DFT spectrum. 1. INTRODUCTION With the advance of mobile devices, people now can easily exchange information via identities in the environment. For example, with a smart-phone camera, one can scan the QR code on a merchandise to get more information about that item. More generally, Davis [1] proposed the concept of signal-rich art which communicates its identity to context-aware devices for pervasive communication. For the example of the QR code, the art is the code and the device is the camera of the smart phone. Inspired by this concept, a method for people to exchange information via a shown on a display device (a TV or monitor) by the use of a smart-phone camera as a capturing tool has been proposed by Wu and Tsai [2]. As shown in Fig. 1, given a message and a, they created a reproduced, called message-rich, for secret message exchange. The message can be extracted in a display-and-imaging fashion i.e., a displayed message-rich is captured with a smart-phone camera to obtain a recorded from which the hidden message is extracted. However, their study of the message-rich is not mature yet; much of the image quality is lost after the message is embedded into the original, as can be seen from the results shown in Wu and Tsai [2]. For example, see Fig. 2, where obvious moiré patterns are seen in the reproduced message-rich. An undesired effect of the moiré pattern on the reproduced, when displayed, will arouse suspicion of third-party persons and reduce the secret-keeping effect of the method. Message: I love you, embedding Message: I love you, message extraction Fig. 1. Illustration of the secret message exchange process by the use of a message-rich proposed by Wu and Tsai [2]. Many kinds of multimedia identities can be utilized to accommodate information for pervasive communication. They can appear in many forms like hard copy, advertisement, poster, display on a monitor or TV, etc. Messages may be injected into such identities by data hiding techniques [3, 4] as mentioned by Davis [1], just like the way information is encoded into the QR code. Such messages hidden in the multimedia identities, which we may call message-rich multimedia, can be extracted by message reading devices such as smart-phone cameras, bar-code readers, etc. using appropriate message extraction techniques implemented by software. There are yet very few studies on the topic of message-rich multimedia. Except the one by Wu and Tsai [2] (Fig. 1) mentioned previously, another is Lee and Tsai [5] which deals with message hiding in images by pixel coding for automatic identification and data capture (AIDC) applications (like QR-code reading), and the reproduced image is called signal-rich-art code image. A block luminance modulation technique for data hiding in the Y component of the YUV color space of the given image and a spatial correction technique for identifying hidden messages are utilized by them to achieve message hiding and extraction. Their method is not applicable to message-rich s. Wen-Hsiang Tsai is also with the Dept. of Information Communication, Asia University, Taichung, Taiwan /16 $ IEEE DOI /ICS

2 (a) Fig. 2. Frames captured from the original and the reproduced message-rich one. (a) A frame from the original. (b) The corresponding frame of the reproduced message-rich generated by Wu and Tsai [2]. From the viewpoint of data recapturing, several data hiding techniques with reproduced images, called stego-images, robust against print-and-scan attacks have been proposed in the past. Here, the print-and-scan operation is equivalent to the display-and-imaging operation investigated in this study. A print-and-scan attack can create distortion of the embedded information after a stego-image is printed and rescanned to form a new one. Chiu and Tsai [6] proposed a method against print-and-scan attacks by embedding coded peak locations in the DFT domain of the input image. For the same purpose, Solanki et al. [7] embedded data into high-magnitude DFT coefficients in low frequencies; and Lefebvre et al. [8] used an additive watermarking method in the spatial domain and a synchronization template in the frequency domain. These methods were proposed for dealing with images, not s. In this study a new method that can produce message-rich s with better image-frame quality and extractable hidden messages is proposed. Problems encountered in the design of the method are analyzed and solved. When a third-party person watches the generated displayed on a monitor, they will have no idea of the existence of the message embedded behind the. This improvement will make the use of the generated message-rich more practical for applications like covert communication, secret message exchanges, etc. 2. SYSTEM DESIGN AND PROCESSES FOR PROPOSED METHOD The hardware system used by the proposed method is shown in Fig. 3, including a monitor or a TV, a smart phone, and a smart-phone holder. Because the to be captured for use by the proposed method might be long, we use the holder to fix the smart phone stably while doing the display-and-imaging work. And the software implementing the proposed method can be decomposed into two parts: a message-embedding process and a message-extraction process. A flowchart of the software processes is shown in Fig. 4. The proposed method can take s with different parameters (like resolution, frame rate, etc.) as the input. And the generated message-rich can survive the display-and-imaging operation. (b) monitor mobile device holder Fig. 3 Hardware configuration of the proposed system. In the message embedding process, at first we get the message to be embedded as well as the frame rate of the input. There is a signal synchronization problem between the message embedding and extraction processes, so we add specially-designed signals to the original message, called initial and ending signals. Then, we transform the new message into a stream of ASCII codes according to the frame rate. Inspired by the method that can resist the print-and-scan attack proposed in [6], the proposed message embedding process modifies the mid-frequency part of the spectrum of each captured image frame by increasing some pixels values in the frequency domain according to the ASCII stream to resist the display-and-imaging operation. Source Embedding Process Message-rich Message Predefined signal for synchonization Capturing the on the monitor Fig. 4 Software processes of the proposed system. Message Extracting process Recorded In the message extracting process, because different smart phones might be used for capturing, the camera parameter setting is taken into consideration at first. Videos with different resolutions might be processed, so the smart-phone camera used to capture the message-rich displayed on a TV or monitor should be set properly so as to capture s with the same aspect ratio. Also, the problem of signal synchronization should be handled. For this, the system is designed to get the frame rate at first. Next, the system analyzes the spectrum of each frame s Y component to extract the initial-signal part of the. Then, the system skips the initial-signal part and starts to analyze the subsequent frames until the ending-signal part is detected. In this process, a noise-removal technique is proposed to correct the extracted embedded ASCII codes to increase message extraction accuracy. 3. MESSAGE EMBEDDING In this section, the propose technique for embedding messages into single image frames is presented at first after describing the weakness of the method proposed by Tsai and Wu [2], followed by the technique proposed for embedding messages in s. (A) EMBEDDING MESSAGE SIGNAL BY PEAK VALUES IN THE FREQUENCY DOMAIN In Wu and Tsai [2], the embedding process injects message signals into each image frame of the input by a way of setting the pixel values to be zero in four pre-selected circular regions in the low-frequency part of the DFT (discrete Fourier transform) spectrum of the Y component of the YUV space of each frame. It is known that the low-frequency part of the DFT spectrum is the most significant portion of an image so that if the number of processed pixels in the low-frequency band is 360

3 too large or/and the changes of these processed pixels values are too great, severe distortion appearing as moiré patterns can be generated as shown in Fig. 5. Such distortion will arouse the suspicion of a third-party person, and reduce the secret communication effect of the proposed method. So we try to find another appropriate way of signal injection in the spectrum in this study for message embedding as described next. (a) (c) Fig. 5 An experimental result yielded by Wu and Tsai [2]. (a) The original image. (b) The Y component of (a). (c) DFT spectrum of (b) with four message signals embedded at a low-frequency part. (d) The image of the inverse DFT of (c) in which moiré patterns can be seen. In the previous method proposed by Chiu and Tsai [6], good results can be obtained by injecting the signals into the mid-frequency part of the Fourier spectrum without creating too much distortion, resulting in an image which is still visually comfortable to human beings. The basic idea of their method is to process only scattered single pixels in the spectrum for message embedding instead of processing pixel regions. Based on the idea of using single pixels in the frequency domain for message embedding, we try to embed signals by injecting peak values into the pixels at the four corners of a high-frequency window centered at the origin of the Y channel s spectrum of each input image frame. The peak values are set to be 5 million at the corners, and the size of the window is a quarter of the image size as indicated by the green rectangle in the figure. After applying a display-and-imaging operation, the embedded signals are reduced by some moiré patterns coming from the resampling effect of the camera s imaging operation, and appear to be very weak in magnitude. Such weak signals could not be extracted to recover the embedded message successfully as shown by our experiments conducted in this study. Modification of the message embedding process should be made. For this, a series of experiments were conducted, in which different window sizes and peak values were tried, and a better result was finally obtained. Specifically, we choose the window size to be a quarter of that of the original image so that the four corners of the window, where peak values for message embedding are injected, fall on a mid-frequency band of the spectrum. And the injected peak values are chosen to be 3 million. These choices of the parameters lead again to a good result of the IDFT which is an image with (b) (d) almost no moiré pattern. Also, after applying the display-and-imaging operation, the peak values in the Y channel s spectrum of the resulting image is still visible. Finally, after the IDFT is applied, not only the distortion in the resulting image is nearly invisible, but the embedded message can also be extracted successfully as will be shown by our experimental results described later. After applying the proposed method described above to a series of images, it was found that the proposed method yields better results than Wu and Tsai [2]. Specifically, the apparent moiré patterns at the edges in the images found in Wu and Tsai s results disappeared after the proposed method is applied, resulting in images which are much cleaner and visually comfortable, as can be seen from Fig. 6. (B) CODING OF MESSAGES FOR EMBEDDING We transform each message to be embedded into a sequence of ASCII codes which is a stream of binary bits, as mentioned previously. Accordingly, we use different arrangements of aforementioned peak values in the spectrum to represent four different binary codes 00, 01, 10, 11 for embedding in a single frame. The arrangement rule is described in Table 1. (a) (c) Fig. 6 A comparison between the result yielded by the method proposed by Wu and Tsai [2] and that by the proposed method. (a) The spectrum of an image into which signals are embedded by Wu and Tsai [2]. (b) The IDFT of (a). (c) The spectrum of the same image in which signals are embedded by the proposed method. (d) The IDFT of (c). Table 1. Design of patterns for message encoding and their meanings used in this study. The binary value represented by each pattern is shown to the right of it. (b) (d) When we embed messages in s instead of in single image frames, we encounter a problem of synchronization, i.e., when we try to record the message-rich, it is almost impossible to capture the displayed right from the starting point of the original. Consequently, it is also difficult to extract signal messages correctly from every frame of the recorded. To overcome this problem, we adopt a synchronization technique which was proposed in [2] by using special signals, called initial and ending signals, based on the property of frames. The initial signal 361

4 is designed to be Accordingly, the message extraction process is started in this study by identifying the real content of the message until it detects two consecutive 3-embedded frames. Similarly, the ending signal is designed to be The message extraction process will be terminated when it detects a triple of 3-embedded frames. Furthermore, smart-phone cameras used by people have different frame rates. If we embed a bit stream into a with a frame rate of 30 fps, for example, another error of synchronization will arise if we record the displayed with a camera with a frame rate other than 30 fps, say, 60 fps. To solve this problem, it observed first that though the frame rate may differ from one another due to uses of different cameras, the content of the message-rich is still the same. For example, if we record a displayed on a monitor with a frame rate different from that of the original, the recorded is still considered to have the same content as the source ; and the content per second is still the same as we watch the. Therefore, the problem can be solved by embedding duplicated signals in a period of time (called embedding period) to achieve the desired effect of synchronization. And the number of times for which each signal is repeated is called a duplicated number. In more detail, if we set the embedding period as X second(s) and the message which we want to embed is S 1S 2S 3, we can inject S 1 into all frames of the first period of X second(s), inject S 2 into all frames into the second period of X second(s), and so on. Then, the extraction process can analyze all frames of every X seconds to retrieve correctly the embedded signals in the recorded. Finally, the details of the proposed message embedding process is described as an algorithm in the following. Algorithm 1: proposed message embedding process. Input: a source V and its frame rate R; a message M; a pre-selected embedding period T; and a pre-selected peak value K for signal embedding. Output: a message-rich in which the message M is embedded. Steps: Step 1. Translate M into a bit stream B of ASCII codes. Step 2. Add initial and ending signals 0033 and 1333 respectively in front of B and at the end of B to form a new stream B new. Step 3. Use the frame rate R of source V and the pre-selected embedding period T to calculate a duplicated number D as D = R T. Step 4. Duplicate each bit of B new for D times sequentially to get a new stream B final. Step 5. Process each frame F of V by the following steps. i. Convert F into the YUV color space. ii. Get the first two bits b 1b 2 of B final as a message signal N and the corresponding peak-value embedding positions P according to Table 1. iii. Remove b 1b 2 from B final to avoid using them again. iv. Embed message signal N into the mid-frequency part of the Y-component spectrum of F by increasing the value of each pixel located at the positions specified by P up to be the pre-selected peak values K. v. If F is not the last frame of V, repeat Step 5; otherwise, continue. Step 6. Convert the final V in the YUV space by the IDFT into the image space as the desired message-rich. 4. MESSAGE EXTRACTION AND CORRECTION While extracting the message from an attacked resulting from a display-and-imaging operation, a problem will be encountered, that is, the change of the peak locations in the recorded frame. Next, different cameras have distinct parameters which can influence the frame rate of the recorded. So, to analyze correctly the frames captured by different cameras, the recorded signals should be segmented precisely according to their respective frame rates. Finally, the signal stream constructed with signals extracted from the frames of a signal-rich might have errors. All of these problems are solved in this study as described in this section. (A) EXTRACTION OF BIT-PAIR SIGNALS FROM SINGLE IMAGE FRAMES When using a camera to record a message-embedded, it is almost impossible to adjust the camera view precisely to match the monitor frame to produce a which has exactly the same size as that of the original one. Therefore, the peak value embedded in each frame s spectrum at a certain position might be moved. Fortunately, due to the DFT property, each peak will not be moved too far away from its original position so that we can set a search region in the spectrum to detect the highest value in the region instead of detecting directly the peak value exactly at its original position. In addition, because the positions where we embed the peak values are just the four corners of a pre-selected window, and due to the symmetry property of the DFT, we have to search only two regions, which are at the left-top and right-top corners of the window, to detect the highest values. We use two threshold values for this search: a magnitude threshold T to decide if each detected highest signal value is high enough to be accepted as a real message signal, and a difference threshold D to decide if the difference between the two signal values detected in the two regions is large enough for verifying whether or not the detected pair of bits is one of 01 or 10 instead of 00 or 11. That is, we have the following decision rule for message signal detection where V 1 and V 2 are the detected highest values in the two aforementioned regions: (1) when the absolute difference between V 1 and V 2 is larger than or equal to D: (1.1) decide signal pattern = 01 if V 1 is larger than or equal to V 2; (1.2) decide signal pattern = 10 if V 1 is smaller than V 2; (2) when the absolute difference between V 1 and V 2 is smaller than D: 362

5 (2.1) decide signal pattern = 11 if V 1 and V 2 are both larger than or equal to T; (2.2) decide signal pattern = 00 if V 1 and V 2 are both smaller than T. (B) EXTRACTION OF MESSAGE SIGNALS FROM VIDEOS In the message extraction process, the initial signal must be detected firstly. Because we use a pre-selected duplicated number to repeat signals in consecutive multiple frames, we cannot detect the initial signal just by a single frame. In addition, the number of frames found in each time period of one second might have small variations for different cases of displaying and capturing; i.e., the frame rate of the recorded is just an average value. If the extraction process uses the average frame rate for signal extraction, a synchronization problem between signal embedding and extraction will arise as well. To overcome these problems, a linear data structure [2] is used to keep the extracted signals so as to segment the signal sequence correctly. For example, let the signal stream be , the embedding period be 1 second, and the frame rate be 4 fps in average. Then, we can get a basic length of 1 4 = 4. When the extraction process detects the first signal pattern 0, it will be saved to a firstly created signal segment with length 1. When the second identical signal pattern 0 is detected, the length of the first segment is extended to 2, and so on; and finally the length of the first segment becomes 3. Next, when the extraction process detects the fourth signal pattern 1, which is different from those in the first segment, a second signal segment is created to keep the signal pattern 1 with length 1. In the end, the extraction process will save four signal segments with each having just two attributes: a signal pattern and a segment length. Comparing with the way of saving all detected signal patterns, this data structure, called signal line subsequently, uses less resources. Also, a signal segment might not be exactly of the basic length. To solve this unequal-length problem, the proposed extraction process analyzes the signal segments sequentially and determines the signal pattern of each segment by comparing the segment length with the basic length. For the above example, the basic length is 4, and the length of the first segment is 3. The extraction process will divide the segment length 3 by the basic length 4 and round off the result 0.75 to get an integer value of 1, ignoring the little difference between them. That is, the first signal segment will be regarded entirely to have a representative signal pattern 0, though the length of this segment is smaller than the basic length with a difference of 1. The aforementioned message extraction process might yield noisy results. For example, a message embedded in a may be detected to be with noise 10. Therefore, a noise removal scheme is proposed to deal with such noise. Conceptually, we treat short signal segments in the signal stream as noise and combine them with neighboring longer segments to construct a better result. In more detail, we consider two cases of a noise segment N in between two signal segments, A and B: (1) A and B have identical signal patterns; (2) A and B have different signal patterns. In Case (1), the signal line is apparently separated by the noise segment N, so that we may combine the two neighboring signal segments A and B with N to construct a larger signal segment with the same signal pattern as those of A and B, and the length of the new segment may be taken to be just the sum of the three shorter segments A, B, and N. In Case (2), if A and B are long enough, we speculate that the noise is on the boundary of the two signal segments A and B, so that the noise N can be combined into either of them. Reversely, if A and B are two short segments, we gather them together and discard the result. Finally, we conduct this noise removal process by an ascending order of their lengths, i.e., we process the shortest noise segment at first. The reason is that we do not want to combine a longer segment with a short one, because it is more possible that longer signal segments are separated by shorter noise ones according to our experimental experience. Accordingly, the above process of message extraction and noise removal is described as an algorithm in the following, where the length of a signal segment L is denoted as L, and the signal pattern of L is denoted as (L). Algorithm 2: proposed message extraction process. Input: a sequence S of signal patterns extracted by the scheme described in Section 4.1, and a basic length l b Output: a message M represented by S. Steps: Step 1. Save S in a linear structure, and transform S into a new one by the following steps. 1.1.If the first signal S i in S is the end of S, go to Step 2; else, create a signal segment L i with its signal pattern as S i and set its length to 1, i.e., set (L i) = S i and L i = Read the next signal S i+1, and if S i+1 is identical to S i, increase L i by 1; else, set S i+1 as the new first signal, increment i, and go to Step Increment i by 1 and repeat Step 1.2. Step 2. Correct each signal segment in S with its length smaller than a half of the basic length l b by the following steps. 2.1 Process the shortest signal segment L N firstly in S by analyzing its two neighboring signal segments L A and L B and combine them by either of the following rules. (A) If (L A) = (L B), then combine all the three signal segments L A, L N, and L B to form a new longer one L ANB and set L ANB = L A + L N + L B and (L ANB) = (L A) or (L B); (B) If (L A) (L B), combine L N with the shorter L S of L A and L B to form a new segment L NS and set L NS = L N + L S and (L NS) = (L S). 2.2 Repeat Step 2 until there is no more short signal segments whose lengths are smaller than a half of the basic length l b. 363

6 Step 3. Convert each signal segment L i in S entirely into a single representative signal by analyzing its signal patterns and length by the following steps. 3.1 Divide L i by the basic length l b, and round off the resulting value to get an integer G. 3.2 Convert L i into a new signal segment whose representative signal pattern is taken to any signal in L i and whose length is set to be G. 3.3 Repeat Step 3 until all line segments are processed. Step 4. Find the initial signal bits from the bit stream S and convert the remaining bit stream into a message by looking up the ASCII charts until the ending signal is detected. 5. EXPERIMENTAL RESULTS We used five advertisement s with different lengths to conduct experiments as shown in Table 2. In the 1st experiment, we computed the rates of detecting low-level binary signals (00~11) in single image frames of the five s using the proposed decision rule for two cases: (1) using only the magnitude threshold T; (2) using both the magnitude threshold T and the difference threshold D. The results are shown in Table 3 from which we can see that the detection rates are increased from Case (1) to Case (2) with an average of 7.30%; and the average detection rate using both thresholds is 98.59%. Though not 100%, improvement can be made by error correction as described in Sec Table 2. Original s used in the experiments. resolution duration 15 seconds 29 seconds 14 seconds 23 seconds 15 seconds fps Table 3. Low-level 2-bit signal detection rates (T: magnitude threshold; D: difference threshold). No. of bianry (1) No. of correctly (2) No. of correctly detected Input No. Message Improvement signals detected using D only using both T and D 1 goodluck % 98.37% 6.50% 2 abcdefghijkl % 99.01% 6.58% 3 iloveyou % 98.08% 9.21% 4 Message! % 99.02% 6.55% 5 NCTU % 98.47% 7.64% average % 98.59% 7.30% Next, we show the rates of extracting embedded high-level message signals, namely, ASCII codes, using the message extraction algorithm (Algorithm 2). We used two embedding periods in the experiments, one being 0.25 second so that an ASCII code can be embedded in every second; and the other being so that an ASCII code can be embedded in every half second. The resulting rates are shown in Table 4 from which we can see that when the embedding speed is higher, the extraction rate is decreased, as expected. For 100% performance, the embedding period of 0.25 second can be used. Table 4. High-level ASCII signal detection rates (T: magnitude threshold; D: difference threshold). Bold and blank characters are erroneously recognition results. Detection using 0.25s embedding period Detection using 0.125s embedding period Input Embdded Extracted No. Rate Embedded message Extracted message Rate message message 1 goodluck88 goodluck88 100% have a nice day hahaha have a jice day jajaha 86.36% 2 abcdefghijkl abcdefghijkl 100% abcdefghijklmnopqrstuvwxyz Abcdefgiijklmno_qrstuvwxyz 92.31% 3 iloveyou iloveyou 100% I love you You love me I_love you Zou love_me 86.36% 4 Message! Message! 100% ABCDEFGHIJKLMNOPQRSTUVWXYZ ABCDEFGHIJKL_N_PQRSTUVWXYZ 92.31% 5 NCTU2015 NCTU % I am tommyfake Y am tommyf_ke 85.71% 6. CONCLUSIONS A new method for creating message-rich s with good quality for pervasive communication like covert communication, secret message exchanges, etc. has been proposed. By transmission of messages via such s, suspicion coming from hackers can be reduced. To generate such s, a new technique has been proposed to embed peak-value codes representing binary ASCII messages into the mid-frequency bands of the Y-component s spectrums of the image frames of the input. The display-and-imaging operation is applied to record the resulting message-rich displayed on a TV or monitor using a smart-phone camera, and message extraction can be conducted successfully from the recorded by detecting high values in the Y-component s spectrums of the frames of the resulting. Several problems about synchronization encountered in the embedding and extraction processes are solved. High message extraction rates are achieved as demonstrated the experimental results, showing the feasibility of the proposed method. Future studies may be directed to improving the proposed method to increase the data amount embeddable in the message-rich, to extract embedded messages in real-time, to resist more types of attacks during display-and-imaging like lighting changes, etc. REFERENCES [1] B. Davis, Signal rich art: enabling the vision of ubiquitous computing, Proceedings of SPIE 7880: Media Watermarking, Security, and Forensics III, N. D. Memon, J. Dittmann, A. M. Alattar and E. J. Delp III, Eds., vol , San Francisco, USA, Feb [2] C. J. Wu and W. H. Tsai, Construction and applications of message-rich s for pervasive communication, Proceedings of 2014 Conference on Computer Vision, Graphics and Image Processing, Pintung, Taiwan, Aug [3] F. A. P. Petitcolas, R. J. Anderson and M. G. Kuhn, Information hiding a survey, Proceedings of IEEE, vol. 87, no. 7, pp , July [4] W. Bender, D. Gruhl, N. Morimoto and A. Lu, Techniques for data hiding, IBM System Journal, vol. 35, no. 3, 4, pp , [5] Y. L. Lee and W. H. Tsai, A new data transfer method via signal-rich-art code images captured by mobile devices, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 25, No. 4, pp , [6] Y. C. Chiu and W. H. Tsai, Copyright protection by watermarking for color images against print-and-scan operations using coding and synchronization of peak locations in discrete Fourier transform domain, Journal of Information Science and Engineering, Vol. 22, No. 3, pp [7] K. Solanki, et al, Print and scan resilient data hiding in images, IEEE Transactions on Information Forensics and Security, Vol. 1, No. 4, pp , [8] F. Lefebvre, A. Gueluy, D. Delannay and B. Macq, A print and scan optimized watermarking scheme, Proceedings of IEEE Fourth Workshop on Multimedia Signal Processing, 2001, Cannes, French, pp , 2001.MapFan.com. (2015, June). MapFan AR Global, [online], available: /arg/ 364