1 Introduction

In tele-medical services, security and privacy of medical information are often highlighted as key concerns [1]. So, it was necessary to develop a method to ensure safety and confidentiality for tele-health services when exchanging patient records via networks. The tele-medical services handle, preserve, and transmit medical records using cutting-edge, cost-effective information technology (IT) systems for a variety of purposes [1]. These services are practical and may be helpful in the medical field, but they also pose a risk of identity theft, privacy invasion, and data tampering at the same time [2]. The modern medical era uses various formats of medical images as per the requirement. Neuroimaging informatics technology initiative (NIfTI) is a format for the storage and representation of medical images. The NIfTI format exists in two different variations: NIfTI-1 and NIfTI-2. The supported data types, accuracy, and voxel size of NIfTI-1 are all improved in NIfTI-2 [10]. Figure 1 shows the comparison between NIfTI and Digital imaging and communications in medicine (DICOM) formats. The copyright protection of these images is the major concern due to their use by the doctors in modern disease diagnosis. The total number of security incidents handled by the Indian computer emergency response team (CERT-In) in 2022 was 1391457 (https://www.cert-in.org.in/s2cMainServlet?pageid=PUBANULREPRT). In which phishing is 1714, probing is 324620, vulnerable services are 875892, malicious codes are 161757, website defacements are 19793, malware propagation is 2164, and others are 5517. These attacks are exponentially increasing yearly, which affects the whole digital ecosystem.

Fig. 1
figure 1

Comparison between DICOM and NIfTI

Full size image

The following are the major novelties of the proposed work:

  • (1) Dual authentication with real-time verification:

The proposed watermarking is highly imperceptible and uses dual authentication. SURF [9] and BRISK [13] feature authentication is used for concerned ROI verification of medical NIfTI images. SURF features are computationally more efficient as well as invariant to blur, scale, rotation, noise, Illumination, and warping [9]. BRISK features are more efficient than SURF [13]. The identity of patient is also verified at the receiver’s end using real time cloud data transfer.

  • (2) High robustness with better quality extraction:

The combination of RDWT [3], RSVD [12], HD [4], and DCT [5] provides high robustness and the quality of the extracted watermark image is also better. RDWT provides a vital advantage over traditional transforms as it is shift invariant [8]. RSVD is used instead of conventional singular value decomposition (SVD), as RSVD has lesser computational complexity than SVD and also requires less memory storage [8]. DCT converts the image into various frequency ranges which further enhances the imperceptibility by selective embedding [5].

  • (3) Optimized scaling factor:

Scaling factor plays an important role to balance the crucial properties of a watermarking system. The proposed watermarking method uses two different optimization technique i.e., PSO [5] and BO [21]. The comparison of these two techniques based on performance parameters are also computed.

  • (4) Enhanced security:

SHA-384 [14] is used to generate the hash value corresponding to MAC address and SHA-512 [14] is used for identity number of patient generated by hospital. Further the Aadhar card of patient is splitted into two parts and previously generated hash values are used to get the hybrid watermark with the help of lifting wavelet transform (LWT) [18, 19] and RSVD.

The rest of the paper is organized as follows: Section 2 contains the literature survey and concerned problems. Preprocessing steps are discussed in Section 3 while Watermark embedding and extraction process is included in Section 4. The simulation results analysis and discussion are mentioned in Section 5. Comparison of results are done in Section 6 whereas conclusion and future works are included in Section 7. Figure 2 shows the organization of paper graphically.

Fig. 2
figure 2

Organization of paper

Full size image

2 Literature survey

In this section, state-of-the-art literature survey is discussed. A novel watermarking technique using Hamming code for channel noise distortion is proposed in [1]. The combination of encryption and compression is used for embedding. Three different compression techniques and their comparison are also mentioned. The work proposed [1] is highly robust and imperceptible but can further be improved for rotation and cropping type of attacks. Hybrid PSO and firefly optimization-based watermarking is proposed in [2]. RDWT and RSVD is used for embedding. Hybrid watermark is generated is generated by using computational inefficient discrete wavelet transform (DWT) and SVD. Encryption technique is used to enhance the security of the proposed work. The proposed work can further be improved against geometric type of attacks.

RDWT-RSVD based image watermarking techniques with enhanced security is proposed in [3]. Hash value and turbo code are used for generation of intermediate watermark. The proposed method tested for various scaling factor. The scaling factor can further be evaluated using optimization techniques and performance can be improved against histogram equalization.

Robust and imperceptible watermarking with HD along with PSO and JAYA optimization (JO) is proposed in [4]. The comparison of different wavelets and optimization techniques are also evaluated. The comparison of LWT and discrete wavelet transform (DWT) is mentioned in [5]. The comparison of PSO and JO are also calculated for different parameters. The proposed method can further be improved against geometric type of attacks.

The comparison of PSO and JO is mentioned in [6]. The block like structures is appeared when the watermark image is extracted against rotation and dither attacks. LWT-DCT-SVD along with Schur decomposition (SD) based DICOM image watermarking technique is proposed in [7]. Firefly optimization technique is used to get the optimum scaling factors and make a proper balance between the vital properties of watermarking. SURF features are used for feature verification. A novel BacterialWmark technique is proposed in [8]. Bacterial foraging optimization technique is used for optimum scaling factor. Homomorphic transform with DWT based robust watermarking technique is proposed in [11]. The robustness can further be improved against geometric type of attacks. NIfTI image watermarking method is proposed in [14] with the combination of LWT and Arnold map (AM). Robustness of this method can be improved further against various attacks. NIfTI image watermarking with LWT, QR decomposition and Multiresolution SVD is proposed in [15].

Singh et al. [16] proposes a NIfTI image watermarking by the utilization of LWT, QR, and multiresolution SVD. Multiple watermarks are used to enhance the copyright protection. The robustness of the proposed method can be further enhanced against some of the attacks. A multifunctional medical image watermarking technique is presented by Sinhal et al. [17] to provide ownership protection, tamper detection (for ROI and other region of non-interest (RONI) segments), and 100% reversible self-recovery of the ROI. Using the Lempel–Ziv-Welch technique, the recovery information of the host image's ROI is first compressed. After that, a transform domain based embedding approach is used to implant the strong watermark into the host image. Additionally, the SHA-256 technique is used to construct the 256-bit hash keys for the robust watermarked image's ROI and eight RONI regions. Using an LSB replacement based fragile watermarking approach, the compressed recovery data and hash keys are merged and then inserted into the segmented RONI area of the robust watermarked image.

An effective watermarking method is presented by Thakur et al. [20] to improve the effectiveness of the DWT-SVD-based strategy. The methodology reduces channel noise distortion and enhances technique security by utilizing well-known error-correcting code and chaotic encryption, respectively. The sub-bands are chosen for the watermark embedding after DWT transforms the cover image. The chosen sub-bands are then subjected to additional transformation via SVD.

Tan et al. [22] proposes a steganography approach-based on generative adversarial networks. The suggested concept consists of three subnetworks: an extractor takes the payload out of stego images, a generator embeds it in cover images, and a potent steganalysis serves as a discriminator to improve steganographic security. By taking use of channel interdependencies, this work constructs a particular channel attention module that dynamically tunes channel-wise characteristics in the deep representation of images. In order to divide the embedding capacity among RGB channels adaptively, Liao et al. [23] present a novel channel-dependent payload division technique based on amplifying channel modification probability. To enhance empirical steganographic security against the identification of channel co-occurrences, the modification probabilities of three corresponding pixels in RGB channels are simultaneously enhanced. This could lead to a clustering of embedding impacts.

Based on image texture attributes, an adaptive payload distribution in multiple images steganography technique is presented in [24]. It also offers a theoretical security analysis from the perspective of the steganalysis. In this work, two payload distribution algorithms are described, one based on image texture complexity and the other on distortion distribution. The suggested tactics can be utilized in conjunction with these cutting-edge single picture steganographic techniques. Liao et al. [25] describe an order forensics framework based on convolutional neural network (CNN) for visual operator chain detection. The purpose of the two-stream CNN architecture is to record evidence of both local noise residuals and tampering artefacts. The novel CNN-based technique, which may automatically learn manipulation detection features straight from image data, is specifically suggested for forensically detecting a chain consisting of two image operators.

An image encryption technique based on bit replacement, chaotic systems, and deoxyribonucleic acid (DNA) coding is presented by Yousif et al. [26]. Its goal is to safeguard the digital photos' privacy and confidentiality. Initially, using this method, every pixel in the image is translated into its matching binary sequence, which consists of bits that are zeros and ones. Subsequently, the one bit is substituted with (0 and 1) bits, and the zero bit is replaced with (1 and 0) bits. By repeatedly altering all of the image pixel bits, two distinct images are ultimately produced. Second, high dimensional chaotic systems based on permutation and diffusion are used to encrypt the resulting images. Finally, DNA encryption is also utilized. An innovative method of securing images is presented in [27] by combining chaotic systems, the El-Gamal public key cryptosystem, and scanning techniques. To put it briefly, a permuted image is initially constructed using spiral and zigzag scanning. Next, the permuted image is encrypted using the El-Gamal encryption technique. Finally, in the confusion and diffusion stages, the pixel locations are randomly generated using Lorenz and Rössler chaotic sequences. A block-based method is put out by Salim et al. [28] to safeguard the authenticity of images by identifying and pinpointing counterfeit. It uses a watermarking technique based on visual cryptography to give forgery detection and localization features. Features, key shares, and secret shares are generated in this watermarking system. DWT, Walsh transform, and local binary pattern are used to extract features from equal-sized blocks of the image and create the feature share. Subsequently, a random key share is generated from every image block, and the secret share is created by XORing the feature share, watermark, and key share.

  1. (1)

    The metaheuristic algorithms help us in computing the optimum scaling factor under certain conditions effectively. If we do not consider these approaches and go for the hit and trial method for finding the scaling factor. This process will not result in a proper trade-off among different characteristics. So, in order to maintain this trade-off, it is required to consider this optimization technique for efficient watermarking.

  2. (2)

    There are various optimization techniques available in the literature out of which we have to choose that provides us the better optimal results under certain conditions.

  3. (3)

    Further, the metaheuristic algorithm is selected in such a way that have less elapsed time with a fast convergence rate.

  4. (4)

    Thus, overview of the existing techniques discussed above leads to the research findings to fill the research gap that motivates us to propose an efficient optimized watermarking technique with optimum gain value using the desired optimization techniques.

3 Preprocessing

In this section hybrid watermark is generated using SHA-384 [14], SHA-512 [14], LWT [18, 19], and RSVD [12].

Following are the steps of hybrid watermark generation:

  • Step 1: Take the Aadhar card of the patient and split it into two parts as shown in Fig. 3.

  • Step 2: Apply one level LWT to both the parts of Aadhar card as shown in Eq. (1):

Fig. 3
figure 3

Two segments of Aadhar card of patient

Full size image
$$\left\{\begin{array}{c}\left[{LL}_{1}, {LH}_{1}, {HL}_{1}, {HH}_{1}\right]=lwt2\left(W1\right)\\ \left[{LL}_{2}, {LH}_{2}, {HL}_{2}, {HH}_{2}\right]=lwt2\left(W2\right)\end{array}\right.$$
(1)

  • Step 3: Apply RSVD to previously generated \({LH}_{1}\), and \({HL}_{2}\) as shown in Eq. (2):

$$\left\{\begin{array}{c}\left[{U}_{1},{D}_{1},{V}_{1}\right]=rsvd\left({LH}_{1}\right)\\ \left[{U}_{2},{D}_{2},{V}_{2}\right]=rsvd\left({HL}_{2}\right)\end{array}\right.$$
(2)

  • Step 4: Take MAC address and generate corresponding hash value (SHA-384) similarly take the patient number generated by hospital and generate corresponding hash value (SHA-512) and generate corresponding image of size 128 × 128 (Ihash1, Ihash2).

MAC address: 84A938D7266C.

Hash values: 206 187 14 18 48 187 45 158 2 117 39 4 161 216 208 15 216 135 106 232 64 59 39 235 178 82 238 202 231 132 214 242 197 170 60 212 87 212 119 100 15 67 156 132 235 129 108 157.

Hospital generated patient number: NAMEAGESEXDISEASECODE.

Hash values: 162 212 97 133 192 8 208 126 206 244 137 170 158 11 255 87 201 98 66 185 97 239 235 82 156 62 24 13 25 242 173 125 67 164 171 168 107 96 209 32 194 158 140 35 34 7 177 88.

  • Step 5: Apply one level LWT to both the previously generated images corresponding to hash values as shown in Eq. (3):

$$\left\{\begin{array}{c}\left[{LL}_{H1}, {LH}_{H1}, {HL}_{H1}, {HH}_{H1}\right]=lwt2\left({I}_{hash1}\right)\\ \left[{LL}_{H2}, {LH}_{H2}, {HL}_{H2}, {HH}_{H2}\right]=lwt2\left({I}_{hash2}\right)\end{array}\right.$$
(3)

  • Step 6: Apply RSVD to previously generated \({LH}_{H1}\), and \({HL}_{H2}\) as shown in Eq. (4):

$$\left\{\begin{array}{c}\left[{U}_{H1},{D}_{H1},{V}_{H1}\right]=rsvd\left({LH}_{H1}\right)\\ \left[{U}_{H2},{D}_{H2},{V}_{H2}\right]=rsvd\left({HL}_{H2}\right)\end{array}\right.$$
(4)

  • Step 7: Add the corresponding dominant components to generate modified dominant component as shown in Eq. (5):

    $$\left\{\begin{array}{c}{D}_{mod1}={D}_{1}+\beta \times {D}_{H1}\\ {D}_{mod2}={D}_{2}+\beta \times {D}_{H2}\end{array}\right.$$
    (5)

    where \(\beta\) is the scaling factor.

  • Step 8: Apply inverse RSVD and LWT to generate hybrid watermarks (HBW1 and HBW2) as shown in Fig. 4. Hybrid watermarks generation process in shown in Fig. 5.

    Algorithm 1
    figure a

    Pre-processing

    Full size image
    Fig. 4
    figure 4

    Hybrid watermarks

    Full size image
    Fig. 5
    figure 5

    Hybrid watermarks generation procedure

    Full size image

4 PBNHWA process

The proposed NIfTI image hybrid watermarking technique uses RDWT [3], HD [4], DCT [5], and RSVD [12] for watermark embedding and extraction. The PSO [5], and BO [21] are utilized to get the optimum scaling factor. SURF [9] and BRISK [13] matching points are used for ROI verification. Following are the steps of the watermarking process:

  • Step 1: Take the NIfTI image as the input and select a slice as shown in Eq. (6):

$$\left\{\begin{array}{c}path="barin.nii"\\ I=niftiread\left(path\right)\\ C=I\left(:,:,s\left(i\right)\right)\end{array}\right.$$
(6)

where \(s\left(i\right)\) is the slice number and 1 \(\le i\le\) 21.

  • Step 2: Apply the RDWT to the previously selected slice of NIfTI image as shown in Eq. (7):

$$\left[{LL}_{I1},{LH}_{I1}, {HL}_{I1},{HH}_{I1}\right]=RDWT\left(C\right)$$
(7)

  • Step 3: Apply HD to \({LH}_{I1}\) and \({HL}_{I1}\) as shown in Eq. (8):

$$\left\{\begin{array}{c}\left[{U}_{1},{T}_{1}\right]=hess\left({LH}_{I1}\right)\\ \left[{U}_{2},{T}_{2}\right]=hess\left({HL}_{I1}\right)\end{array}\right.$$
(8)

  • Step 4: Apply DCT to previously obtained Hessenberg matrices as shown in Eq. (9):

$$\left\{\begin{array}{c}{D}_{c1}=dct2\left({T}_{1}\right)\\ {D}_{c2}=dct2\left({T}_{2}\right)\end{array}\right.$$
(9)

  • Step 5: Select the low and mid frequency components as shown in Eq. (10):

$$\left\{\begin{array}{c}{main}_{1}={D}_{c1}\left(1:\text{128,1}:128\right)\\ {main}_{2}={D}_{c2}\left(1:\text{128,1}:128\right)\end{array}\right.$$
(10)

  • Step 6: Apply RSVD to previously obtained bands of frequency as shown in Eq. (11):

$$\left\{\begin{array}{c}\left[{P}_{1},{Q}_{1},{R}_{1}\right]=RSVD\left({main}_{1}\right)\\ \left[{P}_{2},{Q}_{2},{R}_{2}\right]=RSVD\left({main}_{2}\right)\end{array}\right.$$
(11)

  • Step 7: Take the hybrid watermark images and apply RDWT as shown in Eq. (12):

$$\left\{\begin{array}{c}\left[{LL}_{{HB}_{W1}},{LH}_{{HB}_{W1}},{HL}_{{HB}_{W1}},{HH}_{{HB}_{W1}}\right]=RDWT\left({HB}_{W1}\right)\\ \left[{LL}_{{HB}_{W2}},{LH}_{{HB}_{W2}},{HL}_{{HB}_{W2}},{HH}_{{HB}_{W2}}\right]=RDWT\left({HB}_{W2}\right)\end{array}\right.$$
(12)

  • Step 8: Apply RSVD to previously obtained \({HH}_{{HB}_{W1}}\) and \({HH}_{{HB}_{W2}}\) as shown in Eq. (13):

$$\left\{\begin{array}{c}\left[{U}_{W1},{S}_{W1},{V}_{W1}\right]=RSVD\left({HH}_{{HB}_{W1}}\right)\\ \left[{U}_{W2},{S}_{W2},{V}_{W2}\right]=RSVD\left({HH}_{{HB}_{W2}}\right)\end{array}\right.$$
(13)

  • Step 9: Apply embedding procedure by taking optimized scaling factor (\(\alpha\)) using PSO and BO as shown in Eq. (14):

$$\left\{\begin{array}{c}{M}_{1}={Q}_{1}+\alpha \times {S}_{W1}\\ {M}_{2}={Q}_{2}+\alpha \times {S}_{W2}\end{array}\right.$$
(14)

  • Step 10: Apply inverse RSVD and insert the modified low and mid frequency bands to rest of the bands and take the inverse DCT.

  • Step 11: Apply inverse HD and RDWT to get the watermarked image (Imarked).

  • Step 12: Take the watermarked image and apply RDWT to get the sub-bands as shown in Eq. (15):

$$\left[{LL}_{W},{LH}_{W}, {HL}_{W},{HH}_{W}\right]=RDWT\left({I}_{marked}\right)$$
(15)

  • Step 13: Apply HD to previously obtained \({LH}_{W}\) and \({HL}_{W}\) as shown in Eq. (16):

$$\left\{\begin{array}{c}\left[{U}_{1marked},{T}_{1marked}\right]=hess\left({LH}_{W}\right)\\ \left[{U}_{2marked},{T}_{2marked}\right]=hess\left({HL}_{W}\right)\end{array}\right.$$
(16)

  • Step 14: Apply DCT to previously obtained Hessenberg matrices as shown in Eq. (17):

$$\left\{\begin{array}{c}{D}_{c1marked}=dct2\left({T}_{1marked}\right)\\ {D}_{c2marked}=dct2\left({T}_{2marked}\right)\end{array}\right.$$
(17)

  • Step 15: Select the low and mid frequency components as shown in Eq. (18):

$$\left\{\begin{array}{c}{main}_{1marked}={D}_{c1marked}\left(1:\text{128,1}:128\right)\\ {main}_{2marked}={D}_{c2marked}\left(1:\text{128,1}:128\right)\end{array}\right.$$
(18)

  • Step 16: Apply RSVD to previously obtained bands of frequency as shown in Eq. (19):

$$\left\{\begin{array}{c}\left[{P}_{1M},{Q}_{1M},{R}_{1M}\right]=RSVD\left({main}_{1marked}\right)\\ \left[{P}_{2M},{Q}_{2M},{R}_{2M}\right]=RSVD\left({main}_{2marked}\right)\end{array}\right.$$
(19)

  • Step 17: Apply inverse embedding by taking optimum scaling factor as shown in Eq. (20):

$$\left\{\begin{array}{c}{M}_{1W}=\left({Q}_{1M}-{Q}_{1}\right)/\alpha \\ {M}_{2W}=\left({Q}_{1M}-{Q}_{2}\right)/\alpha \end{array}\right.$$
(20)

  • Step 18: Apply inverse RSVD and RDWT to get the extracted watermarks.

  • Step 19: Combine both the extracted watermark images to get the merged or complete watermark as shown in Fig. 6. Figure 7 shown the watermarking process.

    Fig. 6
    figure 6

    Combined hybrid watermark

    Full size image
    Fig. 7
    figure 7

    Embedding and extraction process

    Full size image
    Algorithm 2
    figure b

    PBNHWA embedding process

    Full size image
    Algorithm 3
    figure c

    PBNHWA extraction process

    Full size image

5 Optimization process

PSO is inspired by swarm foraging and social behavior. Simple ideas and basic operators are used to construct PSO. PSO has a low cost of computation in terms of memory and performance. PSO begins by randomly initializing the population. To investigate the search space, solutions are assigned with randomized velocities. In PSO, a particle is a solution. There are three distinct features of PSO: (1) best fitness of each particle, (2) best fitness of swarm, (3) velocity and position update of each particle [29].

All bats utilize echolocation to gauge distance, and somehow, they also know the difference between background barriers and food/prey. In order to find prey, bats fly erratically at a set frequency, changing wavelength, and loudness. Depending on how close their target is, they can automatically modify the wavelength (or frequency) of their generated pulses and adjust the rate of pulse emission \(\gamma\) in the range [0, 1]. Figure 8 shows the process used in PSO and BO.

Fig. 8
figure 8

Steps used in PSO and BO techniques

Full size image

Following is the fitness equation used in optimization:

$$fitness=\left[\left(SSIM+\frac{1}{PSNR}\right)+\left(\frac{1}{N}\sum_{i=1}^{N}\frac{1}{NCC}\right)\right]\frac{\alpha }{S}$$
(21)

In Eq. (21), SSIM is the structural similarity index measure [5], PSNR is peak signal to noise ratio [5], NCC is normalized correlation coefficient [5], S is the number of slices of NIfTI image, \(\alpha\) is the balancing factor taken as 10, and \(N\) is the number of attacks (16). Flow charts of PSO and BO are shown in Figs. 9 and 10 respectively.

Fig. 9
figure 9

Flow chart of PSO technique

Full size image
Fig. 10
figure 10

Flow chart of BO technique

Full size image

Parameters used in PSO:

Particle size: 5; Cognitive factor (C1): 2; Social Coefficient: 2; Inertia weight: 0.9; Velocity: [0.5,1].

Number of iterations: 21; Elapsed Time: 5.15 s.

Parameters used in BO:

Particle size: 5; loudness: [1, 2]; pulse rate: [0, 1]; \(\varepsilon\): 0.5 (reducing factor for loudness); \(\mu\): 0.9 (pulse rate increasing factor); Number of iterations: 17; Elapsed Time: 4.76 s.

6 Results and discussion

The proposed watermarking method uses NIfTI image [10] as the input host image of size 256 × 256. NIfTI images consists of different slices and behaves as the stack of papers as shown in Fig. 11 and in this work all the 21 slices are used as host image. It comes in two different file extensions:.nii and.nii.gz. while it does not contain any patient information. The hybrid watermark is generated initially then used for watermarking. The size of the hybrid watermark image is 128 × 256. Table 1 shows the performance parameters used in this proposed scheme.

Fig. 11
figure 11

Various slices of input NIfTI image

Full size image
Table 1 Performance parameters
Full size table

6.1 Imperceptibility analysis

For imperceptibility analysis, the proposed method uses PSNR (dB), mean square error (MSE), SSIM, KLD and Jensen JSD. The acceptable value of PSNR is greater than 27 dB without attack [11]. The value of MSE should to tends to zero for higher imperceptibility. The value of SSIM should be close to 1.0000 or greater than 0.9 for better imperceptibility. Lesser the value of KLD [8] and JSD [8] higher is the imperceptibility. Table 2 shows the values of performance parameters for PSO and Table 3 is for BO. In this proposed work, the value of PSNR is greater than 50 dB for PSO and BO in all the slices. The value of MSE is also close to 0 in all the cases. The value of SSIM is greater than 0.9 in all the cases and the values of KLD and JSD are also acceptable for PSO and BO. The value of PSNR is higher in case of BO. The value of SSIM is almost same in PSO and BO except for S5 (higher in case of BO). The values of KLD and JSD are more efficient in case of BO. So, by observing the Tables 2 and 3 it is concluded that the proposed method is highly imperceptible and BO is superior than PSO. Figure 12 shows the Comparison of PSNR (dB) values for different slices and for PSO and BO.

Table 2 Values of performance parameters (PSO-without attack)
Full size table
Table 3 Values of performance parameters (BO-without attack)
Full size table
Fig. 12
figure 12

Comparison of PSNR (dB) values for different slices and for PSO and BO

Full size image

6.2 Robustness analysis

For robustness analysis, the proposed method uses NCC and bit error rate. The value of NCC should be close to 1.0000 without attack. Tables 2 and 3 shows the values of NCC for all the slices and for PSO and BO. In all the cases the NCC value is 1.0000, it shows that the proposed method is highly robust. Figures 13 and 14 shows the extracted watermark images and corresponding NCC values against various attacks for PSO and BO respectively. The NCC values are greater than 0.9 in all the cases. The least NCC value is for sharpening attack (0.9963). The least NCC value is for sharpening attack (0.9963). The value of NCC is higher for BO in cases of Salt and pepper noise (0.01), Histogram equalization, and Motion blur (5, 5). The value of NCC is higher for PSO in cases of Average filter (7 × 7), and Median filter (7 × 7). Rest in all the cases the values of NCC are similar. Table 4 shows the NCC values against various attacks with PSO. Table 5 shows the NCC values against various attacks with BO. Tables 6 and 7 shows the NCC values against hybrid attacks with PSO and BO respectively.

Fig. 13
figure 13

Extracted watermark images and corresponding NCC values against various attacks (S1-PSO)

Full size image
Fig. 14
figure 14

Extracted watermark images and corresponding NCC values against various attacks (S1-BO)

Full size image
Table 4 Value of NCC against various attacks for different input images (PSO)
Full size table
Table 5 Value of NCC against various attacks for different input images (BO)
Full size table
Table 6 Value of NCC against hybrid attacks for different input images (PSO)
Full size table
Table 7 Value of NCC against hybrid attacks for different input images (BO)
Full size table

6.3 Time complexity analysis

In Table 8, the time taken during the process of hybrid watermark generation, watermark embedding and extraction of watermarks are evaluated. The minimum time is taken by S1 (1.7436 s) and the maximum time is for S20 (1.7521 s). The overall timing analysis shows that the proposed method is computationally efficient and also suitable for real-time applications.

Table 8 Time analysis for different input images (in seconds)
Full size table

6.4 Capacity analysis

The capacity of the watermarking technique is evaluated by obtaining the ratio of the number of pixels in the watermark image and the number of pixels in the host image [27]. In the proposed method, RDWT is applied to input host image of size 256 × 256. After applying RDWT the size of the input image remains the same. During the embedding process, the RDWT is applied to the watermark image, so the maximum embeddable watermark size is 256 × 256. Therefore, the capacity of our proposed technique is (256 × 256)/(256 × 256) = 1.0.

6.5 Feature authentication

A distinctive feature point descriptor and local invariant rapid feature point detector are known as SURF. The initial stage of processing assigns an orientation within a sphere centered on the keypoint after determining its location [9]. The SURF detector uses integral images and the determinant of the Hessian matrix as a foundation to increase the speed of feature discovery.

Another binary descriptor is BRISK. A sample strategy is then used in the keypoint's neighborhood after keypoints have been chosen [13]. Short-distance pairs and long-distance pairs are two subgroups of the pairs of pixels surrounding the keypoint. The orientation of the feature point is established by computing local intensity gradients from long distance pairs. This orientation rotates short distance pairs. Short distance pair-wise brightness comparison test data are used to create a BRISK description.

SURF and BRISK features are used to detect that the copyright protection does not harm the ROI regions of brain image. If these regions are distorted or manipulated by the algorithm used then the scheme is inefficient. Table 9 shows the SURF and BRISK feature matching with and without attack. Figures 15 and 16 shows the number matched SURF and BRISK features for different slices (without attack) respectively. Figures 17 and 18 shows number of SURF and BRISK matched against rotation attack respectively. By evaluating the results of matched features, it is clear that the proposed technique does not distort the ROI of medical images (Fig. 19). Table 10 shows the extracted hybrid watermark images.

Table 9 SURF and BRISK feature matching with and without attack
Full size table
Algorithm 4
figure d

SURF and BRISK feature matching

Full size image
Fig. 15
figure 15

Matched SURF features for different slices (Without attack)

Full size image
Fig. 16
figure 16

Matched BRISK features for different slices (Without attack)

Full size image
Fig. 17
figure 17

Matched SURF features for different slices (Rotation-90°)

Full size image
Fig. 18
figure 18

Matched BRISK features for different slices (Rotation-90°)

Full size image
Fig. 19
figure 19

Comparison of number of matched points under rotation attack (SURF and BRISK)

Full size image
Table 10 Extracted hybrid watermark images
Full size table

7 Comparison of results

The comparison of the proposed method with other existing techniques are shown in this section. Table 11 shows the comparison of PSNR (dB) values in which the PSNR values for different schemes are 34.0455 in [1], 45.5891 in [2], 44.9488 in [5], 45.8186 in [7], and 47.12 in [15]. The presented work is having PSNR of 51.6051 dB which is much higher in comparison with other techniques, it depicts that the proposed work is highly imperceptible. Figure 20 shows the graphical comparison of PSNR values. Table 12 shows the comparison of NCC values and it is clear that proposed work is highly robust in comparison with other mentioned techniques. Table 13 shows the time complexity comparison. Table 14 shows the comparison of matched features against Rotation attack. The proposed method authenticates the features of the watermarked image more effectively in comparison with other existing technique.

Table 11 Comparison of PSNR (dB) values for different existing schemes
Full size table
Fig. 20
figure 20

Comparison of PSNR values against different schemes

Full size image
Table 12 Comparison of NCC values with other existing techniques
Full size table
Table 13 Comparison of time complexity
Full size table
Table 14 Comparison of number of matched features against Rotation (90˚)
Full size table

8 Conclusion

PBNHWA technique uses PSO and BO to obtained the optimized scaling factor which maintains the balance between various performance parameters. For embedding RDWT, RSVD, DCT, and HD are used for achieving efficient performance results. The security of the proposed method is enhanced by using hybrid watermarks whereas SURF and BRISK based ROI are verified to demonstrate the superiority of the presented algorithm. The proposed method is tested under various attacks and found to be highly robust. The significant features of medical images are also successfully authenticated with no distortion for telemedicine applications. The proposed watermarking technique can be used for the copyright protection, security enhancement, improved imperceptibility, robustness, features authentication, and identity verification. Although, the proposed technique delivers sufficient robustness and imperceptibility but still there is a scope of improvement. The presented work can also be tested and verified by the hybrid optimization technique under different set of attacks. This task can be considered as the extension of the proposed work for future studies. Deep learning models can also be combined with traditional watermarking approaches to enhance the performance of the proposed work.