Efficient Watermarking Method Based on Maximum Entropy Blocks Selection in Frequency Domain for Color Images

False-positive problem (FPP) is a one of the challenging tasks for the researchers. It authenticates the wrong owner to access the multimedia content. To overcome, the FPP problem, this paper introduces an efficient watermarking method based on the selection of highest entropy blocks. In this method, cover and watermark images are initially shuffled through Arnold transform. Then, the encrypted images are further processed by a 2-level discrete wavelet transform followed by singular value decomposition. The proposed method has been evaluated with geometrical, filtering, noise, and contrast adjustment attacks on the standard image datasets against five recently developed watermarking methods. The simulation results reveal that the proposed method outperforms the existing methods.


I. INTRODUCTION
Security to multimedia content (image, text, audio, video) is a challenging issue globally [1]. The multimedia content can be secured through the frequently employed methods, namely steganography, cryptography, and digital watermarking. However, out of these said methods, both steganography and cryptography methods cannot stop illegal data transfer and distribution over the Internet. As a result, digital watermarking approach mitigates these concerns by including certain information into the cover image as a watermark while maintaining the image quality [2]. The embedding and extraction are two of the most critical phases of any watermarking method. Watermarking methods are grouped into three categories based on the extraction procedure, namely semi-blind, non-blind, and blind. In the semi-blind methods, the watermark and key are needed during the watermark extraction.
The associate editor coordinating the review of this manuscript and approving it for publication was Varuna De Silva .
Non-blind are those methods which require both cover and watermark images. Blind methods require the original key to retrieve watermark images. A robust and reliable watermarking method should possess the following fundamental characteristics [3]: 1) Computational cost: Watermark embedding and extraction should be done with minimal computing effort. This is especially important when dealing with highquality images. While calculating the computational cost, it is essential to consider the time required to embed and retrieve the watermark. A healthy trade-off must be maintained between robustness and computing complexity. 2) False positive rate: False positive issue arises when the embedded watermark is not identical to the retrieved watermark. This property has been primarily utilized for copyright protection and ownership. 3) Imperceptibility: Perceptual transparency is critical for a watermarking system. The observer should not be able to discern a watermark. There should be no detectable artifacts introduced into the original image by the data-embedding process. 4) Robustness: Detection of the digital watermark is possible even after the image has been attacked. The attacks can be in the forms of linear or nonlinear filtering, image enhancement, resizing, and compression to evaluate the robustness. 5) Security: a watermark must remain hidden and unnoticeable to anybody other than the intended recipient.
The watermark should only be available to those who have the proper credentials, and can be often applied using cryptographic keys to meet this security need. A set of private secret keys may only be obtained by a person who is the legitimate owner of the intellectual property image. Generally, watermark might be hidden in either spatial or frequency domains. Histogram shifting (HS), spread spectrum (SS), and least significant bit (LSB) are often used methods in the spatial domain to alter the watermark [4]. It is possible to achieve high imperceptibility by using these domain methods that are computationally efficient and have a large embedding capacity. In addition, watermarking attacks are not resistant to these methods. The watermarks are hidden using the cover image's wavelet coefficients in the frequency domain. These frequency-domain methods include DFT (discrete Fourier transform), DCT (discrete cosine transform), SVD (singular value decomposition), DWT (discrete wavelet transform), and RDWT (redundant DWT) [5]. These methods perform superior to spatial domain methods in terms of robustness [2]. Researchers have studied many watermarking methods based on grayscale and color images [6]- [8]. However, grayscale images contain less information than the color images. Instead of relying on grayscale watermarking methods, researchers are now working on color image watermarking strategies [9].
Watermarking methods based on the DFT transform proposed by Fares et al. [10]. This approach employs two DFT variations, namely FDFT and QDFT, to obscure the watermarks. In an SVD based Watermarking scheme carried out by Shieh et al. [11], both cover and watermark images are initially exposed to chaotic permutation and then both images are partitioned into predefined block sizes. Experimentation has shown that the suggested scheme outperforms the considered method in all the aspects. A DCT-correlation based watermarking method devised by Das et al. [12] hides watermark based on correlation. A DWT-based watermarking algorithm developed by Garg et al. [6] adds PN sequences into the wavelet coefficients. The single-domain strategies outlined in this article cannot deliver acceptable performance. To increase the overall performance, several digital watermarking methods are often combined [13].
A DCT-SVD based watermarking approach for copyright protection devised by Roy et al. [14] conceals scrambled watermark into DCT coefficient blocks. The results of the simulations reveal that the suggested approach provides excellent resilience and high imperceptibility. A method devised by Lai et al. [15] embeds watermarks into the DWT coefficients. Experiments have shown that this method is both reliable and efficient. Huang et al. [16] also suggested a hybrid watermarking method based on DCT-SVD domain. In this method, SVD and DCT are used together to provide great resilience while maintaining imperceptibility. Using a DCT-DWT watermarking system, Abdulrahman et al. [17] developed a copyright-protecting watermarking scheme. Singh et al. [7] presented a hybrid DWT-SVD watermarking method in DCT domain. In this method, 4 × 4 blocks of DCT middle coefficients are employed to hide watermark followed by SVD. This method is impervious to attacks. According to the research, SVD-based watermarking methods are prone to false positives.
Singh et al. [18] developed a watermarking method based on NSCT domain to incorporate a watermark into an image. In this method, the cover image is modified using the RDWT transform to achieve maximum payload capacity, while the AT transform increases security and robustness. In the color model YC b C r , Roy et al. [2] suggested a watermarking method based on RDWT for color images. A scrambled grayscale image is used in this approach to conceal a watermark in the luminance (Y) component. Watermarking attacks cannot compromise the suggested system, as shown by the simulations. Furthermore, Roy et al. [19] also offered an RDWT-DCT based blind watermarking approach. In this scheme, scrambled logos are put into horizontal wavelet coefficients to form a watermarked image. Ernawan et al. [20] devised a blind watermarking system for a grayscale image in RDWT domain. Here, An encrypted watermark image is used to alter the U matrix value of the LL sub-band of the cover image. In contrast, Arnold scrambling is used to generate the encrypted watermark image. The existing watermarking techniques are summarized in Table 1 for better understanding. From the literature, it can be concluded that the above mentioned methods fails to mitigate the FPP problem and computationally intensive [21]. Furthermore, DWT-based watermarking systems offer the advantages of multi-resolution, superior energy compression, and an undetectable visual quality. Therefore, the key contribution of this manuscript is as follows: 1) This paper introduces an efficient, false-positive problem-free based semi-blind watermarking scheme in DWT-SVD domain for color images.
2) The maximum entropy blocks are employed to hide watermark which reduces the computation cost.
3) The false-positive problem issue is mitigated by embedding watermark into the principal components of cover image. The rest of the manuscript is organized as follows: Section 2 examines the watermarking methods associated with this paper. Section 3 covers the proposed method, whereas Section 4 examines the experimental outcomes. Section 5 brings the paper to a conclusion.

A. DISCRETE WAVELET TRANSFORM
DWT uses low pass and high pass filters to split an image into four equal-sized bands. These sub-bands are LL (low frequency), LH (horizontal), HL (vertical), and HH (diagonal), respectively. In DWT, the high pass filter isolates the edges from the cover image, whereas the low pass filter approximates (reproduce the same) the cover image. At decomposition at each level, the low-frequency sub-band (LL) yields the approximation coefficients. The rest of the three sub-bands provide specific information regarding local changes in brightness in the cover image. The low-frequency sub-band (LL) preserves the maximum energy of the cover image. The LH sub-band gives detailed information vertically with the horizontal edge level. The HL sub-band contains detailed information horizontally with the vertical edge level. A watermark can be placed into any sub-band due to the multi-resolution characteristics of DWT. Usually, embedding a watermark in LL sub-band affects the imperceptibility, whereas it improves the robustness. Moreover, embedding watermark information into high-frequency subbands improves the imperceptibility while compromising with robustness. The complete procedure of 2-level DWT is depicted in Figure 1.

B. SINGULAR VALUE DECOMPOSITION
A matrix can be diagonalized symmetrically using SVD transform in linear algebra. A real or complex rectangular There are some advantages of SVD in watermarking application.
1) The singular values (S) define the image's brightness, while the geometric features of the image are represented by the singular vectors (U, V). 2) If S is stable, then modest changes to the image's singular values will not cause large changes in S. 3) Singular values are listed in decreasing order, and many have lesser values than the initial singular value. It can have a slight and non-noticeable effect on image quality by updating or ignoring such small data during the reconstruction step.  2).

III. PROPOSED METHOD
This section covers the proposed method which has two phases, namely embedding and extraction. The proposed method hides a watermark into the maximum entropy blocks (MEB) using DWT-SVD(DS), termed as MEB-DS. The proposed method is feasible for practical applications since its extraction and detection procedures do not use original image, as shown in Equation 16 (the only side information is required). Therefore, the proposed method is semi-blind which authenticates and provides the security to multi-media content. Figure 2 illustrates embedding and extraction phases of the MEB-DS.

A. EMBEDDING PHASE
This section presents the entire embedding procedure in detail. Arnold transform first encrypts both the cover and watermark images to improve overall security. Further, both encrypted images are followed by a 2-level DWT and HH sub-band is opted out for embedding. This sub-band is subsequently separated into 8×8 blocks followed by SVD to insert watermark into principal component. The magnitude of principal components is comparatively higher than the singular components [22]. Therefore, embedding watermark into the maximum entropy blocks of principal components improves the PSNR and NC values as well as reduce computational cost. The embedding method's steps are as follows: 1) The cover (c) and watermark (w) images are of size (M × M ) respectively. 2) Perform AT on both images (c, w) which results in scrambled images (c,w).

B. EXTRACTION PHASE
In extraction phase, an embedded watermark image is extracted by performing a reverse operation of the embedding process on watermarked image. It is divided into four sub-bands by applying 2-level DWT to watermarked images. The sub-band LL is considered, followed by SVD to extract the watermark logo. The extraction procedure is detailed step by step below.  ) as Eq. (14).

IV. EXPERIMENTAL RESULTS AND DISCUSSION
This section critically evaluates the efficacy of the proposed method (MEB-DS) over imperceptibility and robustness. The imperceptibility is evaluated by two matrices, namely PSNR (peak signal to noise ratio) and SSIM (structural similarity index measure), whereas NC (normalized correlation) examines the robustness. A standard dataset having ten RGB color images is used to simulate the performance of the considered methods under different watermarking attacks. The Lena, Barbara, Airplane, Pepper, Baboon, Tulip, Sailboat, Swan, Bear, and Deer are used as cover images, while the Peugeot logo is used as a watermark. Figure 3 depicts both cover and watermark images respectively. The simulation results are carried out on a computer system having 16.0 GB RAM, dual-core 3.7 GHz CPU, and MATLAB 2020a. In terms of imperceptibility and robustness, we have compared the proposed method to five other watermarking methods, including Roy et al. [19], Zhang et al. [23], Liu et al. [24], Bhatti et al. [13], and Ernawan et al. [20].

A. IMPERCEPTIBILITY ANALYSIS
The PSNR and SSIM matrices evaluate imperceptibility between cover and watermarked images. PSNR is calculated by Eq. (19).
where σ c is cover image standard deviation and µ c is the cover image mean. c 1 and c 2 are constants. Both Figures 4, 5, and Table 2 47.40, 0.9998, 0.9996 respectively over the considered cover images, as shown in Table 2. Moreover, Table 3 tabulates the PSNR values, whereas Figure 6 also depicts the PSNR values over the considered methods. The table and figure show that the MEB-DS is better than the existing methods. Hence, the MEB-DS hides watermark logo efficiently without degrading cover image quality.   These attacks are grouped into four categories: Geometrical, filtering, noise and contrast adjustment attacks. The details of these attacks are given. Moreover, each class is further discussed below.       as 0.9959,  0.9864, 0.9866, 0.9914, 0.9910, 0.9865. However, the MEB-DS returns the highest average NC value. It can be observed from both Table 4 and Figure 8 that the MEB-DS performs better than the examined methods in this attack category. 2) Filtering attacks: The three filtering attacks, namely wiener, median, and average with window size (5,5), are encountered on watermarked images to examine robustness in terms of NC. Figure 9 illustrates the watermarked and extracted watermark images against each filtering attack. From the figure, it can be shown that the MEB-DS recovers the superior quality watermark logos. Furthermore, the comparative analysis of NC values returned by considered methods is depicted in both Table 5 and Figure 10. The MEB-DS returns    as 0.9993, 0.9905, 0.9897, 0.9916, 0.9929, 0.9901. The MEB-DS achieves the highest average NC value as shown in Figure 10. Therefore, the MEB-DS performs outstanding against filtering attacks. 3) Noise attacks: The speckle, salt & pepper, and Gaussian noises are incorporated on watermarked image. The quality of the recovered watermark logos corresponds to each noise attack as illustrated in Figure 11. VOLUME 10, 2022  This figure depicts that the MEB-DS recovers a quite recognizable watermark logo under each attack. However, the quality of extracted watermark slightly degraded. Table 6 and Figure 12 compare the NC values over the considered methods. The proposed method returns the highest NC value is 0.9932 against speckle noise. Moreover, the average NC values against considered methods are as 0.9922, 0.9937, 0.9830, 0.9906,  0.9922, 0.9888. These values confirm that the proposed method and Bhatti et al. [13] equally perform over the noise attacks. 4) Contrast adjustment attacks: The sharpening, Gamma correction, and histogram equalization are three contrast adjustment attacks which are encountered on watermarked images. The MEB-DS recovers quite similar watermark logo as embedded watermark logo before, shown in Figure 13. The comparative analysis of NC values over considered methods are tabulated in Table 7 and depicted in Figure 14. The

C. FALSE-POSITIVE PROBLEM ANALYSIS
The reliability and stability of SVD-based watermarking methods make them a popular choice for many applications [15], [25], [26]. Generally, these methods use two approaches to hide the watermark into cover image. The first approach involves inserting the watermark directly into the singular component [15], [25]. Watermark singular matrix values are integrated into a singular component of cover image in the second method. A difficulty  with these domain-based approaches is false-positive problem (FPP) which occurs due to hide watermark in the singular components. This can even result in the incorrect owner being authenticated. To alleviate this issue, the proposed method has hidden watermark logo in the principal component.

D. TIME COMPLEXITY ANALYSIS
This section computes the complexity of the MEB-DS along with existing methods in terms of time. The time complexity is proportional to the time required for embedding and recovering a watermark. The MEB-DS has been compared to five recently existing methods. Liu et al. [24] method uses DWT, SVD, HD, and FOA algorithms, termed as DWT-FOA.
Bhatti et al. [13] method uses Quaternion Fourier transform (QFT), Arnold transform, and Chaotic encryption (CE), termed as (QFT-AC). The proposed method uses only three methods: Arnold Transform, DWT, and SVD. However, the proposed method relies on maximum entropy blocks to hide the watermark. Table 8 and Figure 15 compare all methods over embedding and extraction time. The table demonstrates that the proposed method takes significantly less time to hide and recover the watermark logo than the existing methods. As a result, the proposed method is very efficient due to its low time complexity.

V. CONCLUSION AND FUTURE WORK
Color images are becoming more prevalent in people's lives due to the fast growth of Internet technology, and the need for color image copyright management is becoming increasingly urgent. Therefore, this paper introduced a false-positive free watermarking method based on the selection of maximum entropy blocks in frequency domain. The proposed method improves the imperceptibility by hiding watermarking into the maximum entropy blocks. Moreover, false-positive problem is mitigated by embedding watermark within the prin- ALAKNANDA ASHOK received the Ph.D. degree in digital image processing from the Indian Institute of Technology (IIT), Roorkee. She had experience in various capacities like Nodal Officer of G. B. Pant University of Agriculture and Technology, Pantnagar, twice the Controller of Examination of UTU, Dehradun, a Senior Principal Scientist at CSIR, HQs Delhi, the Director of the Women Institute of Technology, Dehradun, a Committee Member of Ph.D., Academics, Administrative Council, and the Admission Committee of Uttarakhand Technical University (UTU), Dehradun. She has been working as the Dean of the College of Technology, G. B. Pant University of Agriculture and Technology, since February 2020. She has organized several national/international conferences, symposiums, workshops, webinars, and training programs. Several MOUs have been signed under her leadership. She has over 24 years of professional experience. Her work is in digital image processing application of renewable energy and application of the Internet of Things (IoT), wireless sensor networks, smart GIS-based water info systems, and e-health application. She is a member of various academic societies. She had chaired and was the keynote speaker in several conferences.
SUMIT AOLE received the B.E. degree in instrumentation engineering from RTM Nagpur University, in 2012, and the M.E. degree in instrumentation and control engineering from Mumbai University, in 2015. He is currently pursuing the Ph.D. degree with the Department of Instrumentation Engineering, SGGSIE & T, SRTM University, and attached to Universiti Teknologi PETRONAS, Bandar Seri Iskandar, Malaysia, for his research work. He also works as a Project Associate with the CSIR-CSIO, Department of Biomedical Applications, Chandigarh. His research interests include biomedical signal processing, wearable sensor technology, gait analysis, control systems, and rehabilitation robotic devices.
NAVEEN SHARMA is currently working as a Scientist with the Biomedical Division, CSIR-Central Scientific Instruments Organization, Chandigarh. He has ten years' experience in research and development. His research interests include biomedical image processing, computer vision, affective computing, machine learning, and cyber forensics and security. He is a Life Member of the Indian Science Congress. VOLUME 10, 2022