Hybrid Optimization-Based Robust Watermarking Using Denoising Convolutional Neural Network

Colour images have been widely used in many aspects of life; however, copyright violation issues related to these images motivate research efforts. This paper aims to develop an enhanced watermarking algorithm for producing a watermarked image using hybrid optimisation with high imperceptibility and robustness. The algorithm is based on spatial and transform domains and begins by embedding multiple secret marks into cover media using an optimal scaling factor. The multi-type mark contributes an additional level of authenticity to the proposed algorithm. Furthermore, the marked image is encrypted using an improved encryption scheme, and the denoising convolutional neural network (DnCNN) is employed to enhance the robustness of the proposed algorithm. The results reveal that the proposed watermarking algorithm yields low computational overhead, excellent watermark capacity, imperceptibility, and robustness to common ltering attacks. Moreover, the comparison shows that the proposed algorithm outperforms other competing methods.


Introduction
In recent years, internet-based services have been widely used, and colour images are extensively shared online as a common information carrier [1]. Online services such as social media, e-banking, e-healthcare and e-learning have become an important part of our day-to-day activities [2]. This insecure channel has been effectively used for the last few decades for the transmission of multimedia content. Internet technologies have proved to be high speed, indispensable and cost-effective at transmitting media, but, at the same time, they do not provide reliable security when transmitting such content [3]. In these circumstances, transmitting media can be risky and, due to certain security concerns, is an open issue for potential researchers to examine how to make multimedia content more secure. In general, to address these problems, robust and secure watermarking schemes have drawn the attention of the scienti c community. In this scheme, digital marks are invisibly concealed into carrier media to maintain the ownership and integrity of multimedia content [4,5]. A digital mark (s) is a data string that can be used for several purposes. Digital watermarking properties such as robustness, invisibility and embedding capacity are mutually exclusive in nature and, therefore, must be balanced to achieve high performance [6].
Optimisation-based watermarking is frequently used by researchers to manage invisibility and robustness at the same time [7][8][9][10]; however, these approaches still suffer from limited watermark capacity and security.
To overcome the issues discussed above, we have developed an enhanced watermarking algorithm for colour images.
Our contributions are summarised as follows: 1) The fusion of spatial (magic cube) and transform domain (LWT-Schur-T-SVD ) encryption methods are adopted to implement our algorithm, which embeds multiple marks. LWT examines the input in integer form and eliminates the reversibility issue of other wavelet transforms [11]. Furthermore, Schur decomposition allows for speedier and more robust watermarking [12]. The tensor SVD [13], as opposed to the traditional SVD, is used to embed watermarks for improved robustness, and a pseudo-magic cube scheme [14] is used to enhance embedding capacity.
2) Hybrid optimisation procedures called 'HPSOF' [15] have been designed and are employed to compute the scaling factor. A good relationship between invisibility versus robustness is maintained through the factor. 3) To enhance security, improved SIE scrambling [16] is used to encrypt the watermark image. 4) DnCNN is used on the recovered watermark image to improve the scheme's robustness [17]. 5) Multiple marks are concealed in the host media channel to produce the nal watermark. This contributes an additional authenticity to the suggested algorithm.
The remaining sections of this paper are arranged as follows: the literature review is presented in section 2; the proposed hybrid optimisation-based watermarking approach is discussed in section 3; the experimental results are presented in section 4; and the paper concludes in Section 5.

Related Work
In this section, a few well-known related colour image watermarking techniques are brie y discussed, and Table 1 summarises and compares the contributions of various notable methods. Sharma et al. [18] demonstrated a watermarking approach using arti cial intelligence in the transform domain. To provide an additional level of security, the mark is encrypted with chaotic maps. To embed the mark, the singular score of the host media is modi ed with respective RGB channels of principle components of encrypted watermarks. Experiments demonstrate that the scheme is robust and secure. In [19], the author suggested spatial domain-based watermarking using Schur decomposition. This scheme has the advantages of spatial as well as transforms domain techniques, resulting in fewer computations and improved robustness; however, it has a lower embedding capacity. Sharma et al. [20] demonstrated a watermarking approach incorporating the fusion of lifting wavelet transform (LWT) and discrete cosine transform (DCT) techniques and further applied the arti cial bee colony (ABC) algorithm to improve visual quality and robustness. This scheme is robust against various attacks, but a detailed security analysis needs to be conducted. Singh [21] developed a robust watermarking scheme for telehealth applications using LWT and DCT techniques. Before being concealed in the host media, the signature watermark is encrypted with message-digest (MD5) coding, and the patient report is encoded with BCH coding to enhance the robustness and privacy of the scheme; however, the scheme is less robust following a few attacks. Mohan et al. [22] developed a hybrid optimisation-based watermarking approach using an optimisation algorithm that enhances quality and robustness. Here, an additional level of security is provided by applying selective encryption to the host media at a low cost. In [23], the author implemented a spatial domain-based scheme using encryption and direct current coe cients. The mark is divided into different sub-watermarks and is then encrypted using MD5. Furthermore, the encrypted watermarks are concealed in the blue channel components. The presented scheme has improved robustness and the analysis ensures invisibility; however, the embedding capacity needs to be improved. Zear and Singh [24] suggested a hybrid watermarking scheme based on LWT-DCT-SVD to secure multimedia content. The security of the marked image and robustness of the text mark are improved by employing MD5 and Hamming errorcorrecting codes, respectively. The present technique is lacking in terms of BER and optimisation. Kumar and Singh [25] presented a colour image watermark in YCbCr where the embedding and extraction are processed through an alpha blending scheme in the LWT domain. Here, security is enhanced via Arnold's cat map (ACM). This scheme needs to be improved against median lter attacks. Loan et al.
[26] designed a watermarking method that is applicable to both grey and colour images using DCT. The present technique uses a double layer of security by employing chaotic encryption and ACM for watermarked images. From the experimental results, it is noted that the method provides better performance in terms of security, quality and robustness; however, the complexity and cost are high. Haghighi et al. [27] implemented LWT, DCT and a feed-forward neural network-based semi-fragile scheme for temper detection and recovery. Using the inverse halftoning technique, the tempered regions are identi ed in the recovery stage.

Proposed Algorithm
In this section, the proposed algorithm, including computation of the embedding factor, the embedding and extraction phases of multiple marks, and the denoising process of recovered mark data, are described in detail. Figure 1 shows the block diagram of the proposed approach.

Watermark embedding and extraction procedure
In the embedding process, the colour host image, 'cover_img', is initially converted into red, green and blue components represented by 'Rc', 'Bc' and 'Gc', respectively.
This algorithm subsequently utilises DWT to decompose the 'Rc' component into a different sub-band and conceals the PAN number, 'PAN', in the 'HHr' sub-band of the image. Next, the account number, 'Account', is concealed in the green channel, 'Gc', using the magic cube, 'm_cube', algorithm. Following this, the watermark image, 'w_image', is scrambled using the improved SIE scheme and is followed by the T-SVD decomposition. The blue component, 'Bc', is decomposed using the fusion of LWT, Schur and T-SVD. The resultant singular matrix, 'Sb', is altered using the scrambled watermark, 'enc_wimg', and the optimised scaling factor value, 'opt_α'. Lastly, the nal watermarked image is formed by combining all three marked channels (wat_R, wat_G, wat_B). The watermark extraction procedure is the inverse of the embedding process. The complete embedding and extraction processes are explained in Algorithm 1 and Algorithm 2, respectively. Table 2 lists the speci cs of the notations used in the algorithms.
Table2. Description of the notations used in the algorithms

Determination of optimal embedding factor value
To determine the ideal optimal embedding factor value for the embedding of multiple watermarks, the HPSOF optimisation scheme is used, which provides a well-balanced trade-off between invisibility and robustness. PSO and Fire y are swarm intelligence-based metaheuristic methods inspired by nature. To minimise the cost of complex numerical problems, Aydilek used a fusion of PSO and Fire y [15] optimisation techniques. Table 3 includes the parameters used for controlling the algorithms to produce optimal outcomes. Also, the tness value objective function is de ned as: 1 Where, ' ' is a stabilizing factor which balances the quality and robustness effects. Algorithm 3 summarizes the complete process of determining the best scaling factor ('opt_sfv') for watermark embedding.  [29] were used and are shown in Figure 2.
Each test was performed on colour host images of 512×512 pixels, user PAN and account details of 80 and 96 bits, respectively, and a grey mark image of 128×128 pixels. The mark details are presented in Figure 3. To evaluate and test the proposed algorithm, we have carried out an invisibility analysis using PSNR and SSIM [2], a robustness analysis using NC and BER [2] and a differential analysis using NPCR and UACI [17].
The invisibility, robustness and security performance of the suggested method on the Kodak and USI-SIPI datasets are shown in Table 4. It is noted that the average PSNR value obtained for 40 test images is 57.7124 dB, with the highest value among them being 59.5111 dB. Additionally, the value of NC and BER is approaching 1 and 0, respectively, in all cases. The average NPCR reaches 0.9956, and the average UACI score reaches 0.2747, demonstrating the encryption scheme's high security capabilities. Therefore, the results shown in this table are indicative of the suggested method's positive potential.  Figure 4 and Table 5, respectively, for 40 test images using the Kodak and USI-SIPI datasets. From Figure 4, it is noted that the satisfactory score of NC, i.e. NC ≥ 0.7863, is achieved for all considered attacks except the cropping attack. Similarly, the average BER is shown in Table 5 for the same set of dataset images. The average BER score achieved is zero, except for a few attacks (i.e. median ltering, resize and cropping).  Table 6 and Figure 5, respectively, and, compared with several other related methods, it can be observed from Table 6 that the achieved PSNR score outperforms the recent methods. Similarly, from Figure 5, it can be noted that our average NC performs well apart from the cropping attack.

Conclusions
A secure colour image watermarking algorithm with improved invisibility, robustness and capacity has been proposed in this paper. In the suggested algorithm, embedding is performed in the spatial and transform domains. Initially, the host colour image is divided into red, green and blue channels. Then, multiple-mark information is concealed in all three channels for different purposes. A trade-off between invisibility and robustness was analysed by fusing PSO and Fire y algorithms. To evaluate the performance of the proposed work, invisibility, robustness and security analyses were performed against various attacks. A detailed comparison was also made, demonstrating how the proposed research is better than the existing algorithms. In the future, we would like to explore the concepts of encryption and compressionbased watermarking, thereby making our system more e cient for other media applications.

Declarations
Authors' contribution Both authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by both authors. The rst draft of the manuscript was written by Dhiran Kumar Mahto and Amit Kumar Singh commented on previous versions of the manuscript. Both authors read and approved the nal manuscript.
Funding This work has no funding resource.

Declarations
Con ict of interest The authors declare that they have no con ict of interest.
Ethical approval This article does not contain any studies with human participants or animals performed by any of the authors.
Consent of publication Not applicable. Used some of the test dataset Host images Used watermarks