Design of artwork resource management system based on block classification coding and bit plane rearrangement

Block classification code

Block classification coding (BCC) is a technique that involves dividing image data into smaller blocks and then classifying and encoding these blocks based on their characteristics (such as color, texture, etc.). The goal is to enhance the efficiency and accuracy of data processing. Conventional block classification coding predominantly involves segmenting images, extracting block-specific feature information, and generating a recovered pattern watermark of either fixed or variable length, thereby achieving a balance between small pattern watermark capacity and high recovery quality (Farri & Ayubi, 2023). In literature (Wu & Yu, 2019), the image block is subdivided into 4 × 4 dimensions, and vector quantization (VQ) (Edoardo Daniele, 2018) compression coding is applied to encode eight bits of fixed-length watermark information for each VQ index of the image block. Meanwhile, literature (Hussan et al., 2022) employs the pixel mean for encoding six bits of information, which is resilient against constant mean attacks, coupled with two bits of watermarking information to pinpoint tampering locations precisely. Furthermore, the literature (Bhalerao, Ansari & Kumar, 2023) employs an additional authentication watermark for tampering localization and determining the position of the tampered recovered image block. In a similar vein, literature (Yoo, Lee & Kwak, 2018) divides the image into 8 × 8-sized blocks, applying the DCT method for processing and subjecting the entirety of the recovered data from each image block to run-length encoding (RLF) to generate the final recovered watermark information.

Hence, to minimize the recovery watermark’s capacity, researchers have proposed the adaptive generation of a variable-length recovery watermark based on the content of the image block. In literature (Dihin, Shemmary & Al-Jawher, 2023), the adaptive embedding objective is achieved by dividing the 8 × 8-sized image block into 4 × 4 or 2 × 2 dimensions, aligning with the texture structure of the block during watermark information embedding. Meanwhile, literature (Dong et al., 2018) segments the image into 4 × 4-sized blocks, discerns the textural or smooth nature of an image block by computing the variance magnitude, and encodes textured blocks and flat sliders with varying lengths. In a similar vein, literature (Sebai, Zouaoui & Ghorbel, 2024) initiates SPIHT coding of the image under a specific compression ratio (Yang et al., 2023), followed by RS error correction coding of the coded image (An, Médard & Duffy, 2021). The addition of authentication watermark information serves a dual purpose: tampering localization and an expanded scope for tampering restoration. This approach addresses both watermark capacity and image restoration quality concerns concurrently. The applicability of this concept extends to image restoration as well. In the quest to enhance the quality of natural image restoration, literature (Press et al., 2020) employs RS coding of the image at a defined compression ratio (Hussan et al., 2023). The incorporation of authentication watermark information facilitates the detection of tampering, thereby enhancing tampering detection and achieving a balance between watermark capacity and image restoration quality. Given the rich array of color, shape, and other image information inherent in art images, authenticating image information and facilitating tampering recovery emerge as pivotal aspects in art resource management.

Bit plane rearrangement

Bit plane rearrangement (BPR) is an image processing technique that decomposes each pixel value within an image into multiple bit planes and rearranges these bit planes according to specific rules. Within art resource management systems, this technique was employed to compress and optimize the image data of artworks. Research demonstrated the superior utilization of pixel correlation through BPR. For instance, in literature (Yao et al., 2023), four rearrangement methods were devised to cluster highly correlated pixel values, consolidating space through run-length code (Yao et al., 2023) compression. Addressing the minimal redundancy in binary images, literature (Chhajed & Garg, 2023) classified them into uniform and non-uniform blocks, embedding secret data strategically. Building on this approach, the literature (Alqahtani & Masmoudi, 2023) enhanced the method by incorporating pixel prediction and bit plane rearrangement, dividing bit planes into uniform and non-uniform blocks to increase embeddable space.

Subsequently, researchers explored the synergy of bit plane rearrangement with pixel prediction. Literature (Alqahtani & Masmoudi, 2023) proposed utilizing the most effective bit plane for information storage, predicting the highest bit plane, marking inaccurately predicted positions, and storing secret information in correctly predicted positions, followed by block rearrangement. Literature (Fu et al., 2024) refined this scheme by introducing bit plane embedding information, thereby augmenting the embedding capacity. Introducing the RDHEI algorithm for binary images, literature (Pang et al., 2023) categorized blocks into uniform and non-uniform based on pixel values, using bit substitution for uniform blocks and pixel prediction for non-uniform blocks. An alternative approach was presented in the literature (Alqahtani & Masmoudi, 2023), suggesting extended tour coding, where scanning the most significant bit plane yielded consecutive identical long bit streams, efficiently compressing the bit plane and preserving space.

Furthermore, literature (Chen & Chang, 2019) enhanced compression by focusing on the prediction error image, resulting in longer consecutive bit streams and improved embedding capacity. Despite these advancements, existing techniques exhibited coding and noise reduction shortcomings. Ongoing research focused on refining compression efficacy and embedding capacity through further advancements in bit plane rearrangement.

RS-BCC algorithm

In this subsection, the RS-BCC algorithm is introduced, leveraging the inherent characteristics of artwork image resources and concurrently addressing watermarking capacity and image recovery quality. The initial steps involve classifying the image into blocks. Red-green-blue (RGB), a widely adopted scheme for recording and displaying color images, represents each pixel of a color image with R, G, and B components. In RGB mode, the component values range from 0 to 255, with an image displaying as pure white when all three components approach 255 and as pure black when they approach 0. Categorization of artwork images is accomplished based on the relative magnitudes of the R, G, and B components, with an emphasis on distinguishing reddish images when the R component exceeds the G and B components. The regions displaying red color due to the presence of a red stamp in the artwork canonical collection are denoted as stamp blocks.

Image blocks containing colors conducive to forming graphics are termed significant blocks, while white-blank regions are designated as blank blocks. Regions displaying other formats are labeled as format blocks. An artwork image I is systematically divided into n non-overlapping 8 × 8 image blocks, each block sequentially numbered for subsequent processing:

(1) $$I_i=\left\{r_{i,j},g_{i,j},b_{i,j}\right\}$$where i = 1, 2, …, n; j = 1, 2, …, 64 denotes the R, G and B components of the jth pixel of the ith image block, respectively. The choice of an 8 × 8 block size is based on a comprehensive consideration of both image processing efficiency and watermark embedding effectiveness. This size not only preserves the image details but also provides sufficient space for efficient watermark embedding and encoding, aligning with the commonly adopted standards in the current field of image processing. Determine the type of the image block $I_i\left(i=1,2,...,n\right)$ and get the block type labeling matrix:

(2) $$O=\{o_i|i=1,2,\dots ,n\}.$$

Counting the number of blank blocks $n_1$, stamp blocks $n_2$, format blocks $n_3$ and significant blocks $n_4$.

First, initializing the block labeling zero matrix $O=\{o_i|i=1,2,\dots ,n\}$. For each image block, the value of pixels displayed as pure white areas is close to 255. Therefore, if the minimum value of the three components of each pixel in the block is greater than 250 and contains no other information, the image block is considered to be a blank block. The value of the block-type label $o_i$ is set to 1.

Significant blocks within the image, distinguished by colors that can form graphics, are particularly susceptible to attacks. The disparity among the three RGB components is relatively minor, except for pixels harboring red text information, where the R component values significantly surpass those of the G and B components. For the image block $I_i\left(o_i\ne 1\right)$ that is not labeled as a blank block, there exists a pixel $r_{i,j},g_{i,j},b_{i,j}$ in it that satisfies the following conditions:

(3) $$r_{i,j}-b_{i,j}\le 10$$

(4) $$b_{i,j}\le 250$$where $r_{i,j},g_{i,j},b_{i,j}$ respectively represent the values of the red (R), green (G), and blue (B) color components of the j^th pixel in the i^th image block. At the same time, the use of technology can be recognized by the regular shape of the graphic; therefore, the image block is considered an essential component. The value of the block type label is set to 2, enabling RS-BCC to accurately distinguish between essential blocks.

The seal block and format block can obtain a binary image using the following inequality:

(5) $$g_{i,j}-b_{i,j}\le 10.$$

According to the above inequality can distinguish the stamp block and format block in the red message. For the unmarked image block $I_i$, first count the number of results of $abc\left(g_{i,j}-b_{i,j}\right)\le 10$ and choose the appropriate threshold value $t_1$. If the counting result is not less than, the image block is considered a stamp block, the value of the block type label is set to 3, and the remaining image block becomes a format block. As the stamp block and format block both manifest as reddish areas, the size of the three component values for some pixels tends to be similar, leading to a less precise differentiation compared to the blank block and the significant block.

In image processing, it is crucial to distinguish among four types of blocks: stamp, significant, format, and blank. Stamp blocks typically feature rich and varied colors, similar to the diverse colors of stamp images. This can be determined by calculating the number of colors and their distribution. If the number of colors exceeds a certain threshold and the distribution is scattered, it is likely a stamp block. Significant blocks contain essential information, such as faces and text. Algorithms for face detection and text recognition can be utilized to assist in judgment while also considering color contrast. Areas with high contrast are more likely to be significant blocks. Format blocks are fixed areas related to the image format, such as borders, headers, and footers. They can be identified through positional information, such as being close to the edges of the image, as well as by their uniform color. Blank blocks are areas that are not filled with effective information where pixel values are similar. They can be determined by calculating the variance of pixel values; when the variance is below a specific threshold, the block is classified as a blank block.

The ultimate block-type labeling matrix O is derived using the following formula:

(6) $$h_i=\{\begin{array}{c}1,{\cap }_{j=1}^{64}(min(r_{i,j},g_{i,j},b_{i,j}))>250\hfill \\ 2,\mathrm{\exists }{r}_{i,j}-b_{i,j}\le 10,r_{i,j}\le 250,b_{i,j}\le 250\hfill \\ 3,sun(abs\left(g_{i,j}-b_{i,j}\right)<10)>t_i\frac{3}{4}\frac{3}{4}and\frac{3}{4}\frac{3}{4}{o}_i=0\hfill \\ 0,else.\hfill \end{array}$$

In the RS coding process, a generator polynomial is used to produce a specific number of redundant parity codes for each data block. These parity codes are then appended to the data block. While this approach increases storage overhead, it significantly enhances the error-correction capability of the data during transmission or storage. Consequently, it indirectly improves image quality, particularly enabling the restoration of the original image when the data is damaged.

Artwork image watermark embedding based on RS-BCC

This section extends the RS-BCC algorithm established in “RS-BCC algorithm” to enhance the security of artwork images. Initially, the original artwork image I is transformed into the YUV color space, and watermark generation and embedding are executed on the Y component. As illustrated in Fig. 1, the binary encoding of the image block type matrix O begins by sequentially generating binary row vectors of length 2n. Subsequently, the disarranged vectors undergo encryption with a key encryption method, and the encoded binary row vectors are subjected to RS(1, 3) encoding. Finally, the encoded binary row vector undergoes further encryption with key scrambling, resulting in the generation of a type-coded row vector:

(7) $$o=\{q_i|i=1,2,\dots ,6n\}.$$

The type code is uniformly embedded as a partially recovered watermark into the lowest bit of pixels 1 to 6 of the Y-component of each image block. In the case of a non-blank block, 64 bits of binary watermark information are generated for each image block:

(8) $$w_{i,j}=\{\begin{array}{c}1,x_{i,j}<tp\\ 0,else\hfill \end{array}$$where t is the threshold value and p is the number of blank blocks. $w_{i,j}$ is aligned into a row matrix $W_i$ with size (1,64*(n − n₁)). The matrix is obtained through key-based disambiguation encryption of the row matrix. The disambiguated matrix is then encoded with RS(2,3), and the resulting encoded matrix undergoes disambiguation and encryption based on the key to yield the binary recovery watermarking matrix with a specified size $(1,96\ast \left(n-n_1\right))$.

Next, according to the number of non-blank blocks, the number of stamp blocks, format blocks and blank blocks to be embedded in the binary recovery watermark is selected, and $n_2$ stamp blocks, $n_3$ format blocks and $\lceil 96\ast \left(n,n_1\right)/58\rceil -\left(n_2+n_3\right)$ blank blocks are selected for embedding ( $\lceil \rceil$ means upward rounding), and the information of the watermark is embedded in the 7th to 64th pixels of the selected image blocks.

Finally, the Y-component of the embedded watermark is converted to RGB space with the U-component and V-component to get the watermarked image $I^w$. To validate the robustness of Bit-Plane Rearrangement (BPR), we conducted compression experiments on images under varying noise levels and compared the results with traditional methods. The outcomes demonstrate that, despite the increased complexity of the images due to noise, BPR is capable of maintaining a high compression efficiency while preserving relatively good image quality. This evidence underscores the effectiveness of BPR under noisy conditions.

BPR

To further mitigate the wastage of space resources in artwork resource management, we undertake BPR and compression on artwork images. Initially, the artwork image is segmented into eight-bit planes, with each bit plane further divided into n non-overlapping blocks, each having a size of S × S. Here, a block composed entirely of “0” is defined as a homogeneous block (HB), while all other blocks are categorized as non-homogeneous blocks (NB). Moreover, using “1” to mark HB and “0” to mark NB, all labels can form a label map (LM) of size (M/S) × (N/S), assuming that the number of HBs is $n_{hb}$, as shown in Fig. 2, which is set up with different coding structures according to the texture distribution characteristics in the bit plane. During the BPR process, we divide the image’s bit-planes into HBs and NBs. Homogeneous blocks are those with relatively consistent pixel values, while non-homogeneous blocks contain more details and variations. In noisy or highly textured images, the proportion of non-homogeneous blocks tends to increase, which may impact compression efficiency. Nevertheless, by optimizing the encoding strategy, such as employing more efficient encoding algorithms or adjusting block sizes, BPR can still achieve effective compression in these types of images.

If $4n_{hb}\le \left(M/S\right)\left(N/S\right)$, the number of uniform blocks in the bit plane is small, and the space in the uniform blocks cannot accommodate the LM, which cannot achieve the effect of compression, and the bit plane is marked as 0. If $4n_{hb}>\left(M/S\right)\left(N/S\right)$ the bit plane can reserve space for embedding the secret information, it needs to be further classified according to the LM. Since the high-bit plane exhibits a strong correlation, the non-uniform blocks in this plane account for only a very small portion of the generated LM with sparse characteristics. Consequently, the LM can be effectively compressed by using extended tour coding. The compressed LM is denoted as LM′, assuming that the length of LM′ is 1. To facilitate subsequent decoding, the length information is recorded using the bit. If $1+Ib\left(M/S\right)\left(N/S\right)\ge \left(M/S\right)\left(N/S\right)$, then the LM is not compressed and the bit plane is labeled as 10; otherwise, it is labeled as 11.

The compressed bit plane undergoes rearrangement to allocate reserved space in the low-bit plane. The bit-plane rearrangement process is illustrated in Fig. 3, where the auxiliary information A is initially stored in the highest bit plane, encompassing the block size S (four bits), the length of the fixed-length encoding $L_{fix}$ (four bit), the original pixel of the first pixel in the upper left corner (eight bit), the number of overflow dots $n_{over}(\lceil IbMN\rceil bit)$ and the location of each overflow dot $a{d}_{over}(\lceil IbMN\rceil bit)$, followed by storing the encoding of the individual bit planes in turn, and the remaining space is used for the storage of the secret information. Thus, the net embedding capacity C of the image is:

(9) $$C=8MN-len\left(A\right)-{\sum }_{i=1}^{8}len\left(P_i\right)-\lceil Ib8MN\rceil$$where len(*) is the length of the message, $P_i\left(i=1,2,3,\dots ,8\right)$ is the encoding of the ith bit plane, and 8 is the highest bit plane. To facilitate the embedding of information by the information embedder, the net embedding capacity C of the image must be stored $\lceil Ib8MN\rceil bit$ below the lowest bit plane.

Compared with the traditional PR technique, the BCC-BPR-ICWE algorithm proposed in this article achieves more efficient image compression and watermark embedding by integrating block classification coding. Especially when processing artwork images, it can better preserve image details and color information.

Evaluation indicators

To assess the image quality post-fast classification-based coding and bit-plane remapping, we employ two metrics: Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).

PSNR serves as a metric for evaluating image quality, quantifying the error between corresponding pixel points in decibels (dB). Given an original image $I_{original}$ and a distorted image $I_{distorted}$ its calculation formula is as follows:

(10) $$\mathrm{P}\mathrm{S}\mathrm{N}\mathrm{R}=20\times {\log }_{10}(MA{X}_I)-10\times {\log }_{10}(MSE)$$where $MA{X}_I$ is the maximum possible pixel value of the image. $MSE$ is the mean squared error between the original image and the distorted image, calculated as:

(11) $$MSE={\displaystyle \frac{{\sum }_i^{m-1}{\sum }_j^{n-1}{({I_{original}}_{\left(i,j\right)}-{I_{distored}}_{\left(i,j\right)})}^2}{mn}}$$where m and n are the height and width of the image, respectively, ${I_{original}}_{\left(i,j\right)}$ and $I_{distorted\left(i,j\right)}$ are the pixel values of the original and distorted images at position (i, j), respectively.

SSIM measures image similarity in terms of brightness, contrast, and structure, with values ranging from 0 to 1, where larger values indicate less image distortion. Let luminance $I\left(x,y\right)$, contrast $C\left(x,y\right)$ and structure $S\left(x,y\right)$. The final SSIM index is the product of these three components and the importance of each component is adjusted by a weighting factor $\alpha ,\beta ,\gamma$ (usually set to 1).

(12) $$SSIM=I(x,y{)}^2+C(x,y{)}^2+S(x,y{)}^2.$$

Model performance comparison

The complexity analysis of the BPR algorithm is of great significance for evaluating its performance. In terms of time complexity, during the encoding phase, assuming the size of an image block is m × n, the time complexity of performing a transformation (such as the DCT transformation) on the image block is O(m × n × log (m × n)). The time complexity of the quantization operation is O(m × n), and the time complexity of entropy encoding is approximately O(k) (where k is related to the length of the encoded data). Overall, the total time complexity of the encoding phase is O(m × n × log (m × n)). The decoding phase is the inverse process of encoding, and its time complexity is also O(m × n × log (m × n)). Regarding space complexity, both the encoding and decoding phases require storing the original image block, intermediate results, and so on, resulting in a total space complexity of O(m × n). Through such complexity analysis, we can clearly understand the resource requirements of the BPR algorithm for data of different scales, providing a basis for performance evaluation in its practical applications.

Figure 4 illustrates the PSNR and SSIM values for various algorithms in generating watermarked artwork-containing images, allowing for a quantitative comparison of algorithmic invisibility. As depicted in Fig. 4, the BCC-BPR-ICWE algorithm presented in this article yields watermarked artwork images with a PSNR exceeding 57 dB. This is attributed to the algorithm’s watermark generation, which considers both tampering authentication and image recovery. Specifically, the algorithm encodes only the text information in the artwork, resulting in a small watermark capacity. The lowest valid bits of some pixels in the watermark-containing image component are embedded in the watermark information, while other pixels remain unaltered.

In contrast, DCT_RLF requires the generation of additional authentication watermarks, which embed watermark information in each pixel. SPIHT_RS and RDHEI embed 2 bits per pixel, resulting in an image containing watermarked artifacts with a PSNR of approximately 42 dB. This corresponds to a watermark embedding capacity of 2 bits per pixel (bpp). Additionally, the watermark capacity embedded in each pixel of SPIHT_RS and RDHEI can reach up to 3–4 bits per pixel (bpp).

The watermark embedding capacity of BCC-BPR-ICWE (Ours) is determined by the number of non-blank blocks and the type code watermarking information, ranging from 0.45 to 0.50 bpp. This is comparatively lower than the relevant literature. Furthermore, according to the SSIM metrics, the BCC-BPR-ICWE (ours) achieves a structure similarity of 0.9993 after watermarking, whereas DCT_RLF, SPIHT_RS, and RDHEI achieve 0.9783, 0.9846, and 0.9417, respectively. This suggests significant structural modifications in the latter algorithms, impacting the artwork resource management system. The BCC-BPR-ICWE (ours) preserves more information of the original image after watermarking the artwork, facilitating easier retrieval of the artwork image containing the watermark.

Figure 5 presents a performance comparison of different algorithms in generating watermarked artwork images. Key metrics include watermark capacity, which indicates the quantity of embedded watermark information, and file size, reflecting the increase in size compared to the original image. This increment is crucial for managing storage overhead in the artwork resource management system.

In Fig. 5, the BCC-BPR-ICWE (ours) outperforms other comparative methods by generating watermark-containing images with a significantly lower file increment. Based on the experimental results, we observed that when the compression rate ranges between 0.5 and 0.7, both image quality and watermark robustness remain at high levels. Simultaneously, the watermark embedding capacity, W, stabilizes around 0.5 bits per pixel (bpp), effectively meeting the requirements for watermark embedding while preserving image quality. This outcome validates the effectiveness of our proposed algorithm in striking a balance between compression and watermark capacity.

Simultaneously, the file size is a mere 154 kB, deemed acceptable in practical applications, as this increment imposes a minimal burden on image storage and transmission. This advantage of BCC-BPR-ICWE (ours) becomes more evident when compared with other comparative literature.

Specific traditional watermarking algorithms may struggle to strike a balance between watermarking capacity and file increment, often resulting in either insufficient watermarking capacity or excessive file increment, both of which are unfavorable for practical applications. However, the proposed BCC-BPR-ICWE achieves a commendable equilibrium between watermark capacity and file increment through optimized algorithm design and parameter adjustments.

Comparison of artwork image tampering and recovery performance

We conducted experiments on image information tampering and recovery using an image from the WikiArt dataset. In the Art Resource Management System (ARMS), an official stamp indicating copyright is typically added to the image. To simulate this, we appended the word “copyright” to the base of the image, mimicking an official seal. This setup allows us to demonstrate block classification in block categorization coding and assess structural preservation performance in image recovery.

Figure 6 presents the tampering detection results of BCC-BPR-ICWE, DCT_RLF, SPIHT_RS, and RDHEI, which are proposed in this article. Sub-figure (1) in Fig. 6 represents the tampered image. Figures 6A, 6B, and 6C display the recovery results obtained using the algorithms proposed in this article, namely DCT_RLF and SPIHT_RS, respectively. Figure 6D illustrates the RDHEI algorithm with the same recovery watermarking capacity. The RDHEI algorithm yields the poorest recovery results. Through detection, we determined its recovery watermark capacity to be 1.2 bits per pixel (bpp).

Conversely, Fig. 6D obtained by the BCC-BPR-ICWE algorithm in this article exhibits the strongest structural recovery.

As depicted in Fig. 7, the PSNR value for the artwork image recovered by the DCT_RLF algorithm is 23.97 dB with an SSIM of 0.9922. The SPIHT_RS algorithm achieves a PSNR value of 36.87 dB and an SSIM of 0.9902. Notably, the image recovered by the BCC-BPR-ICWE algorithm attains a PSNR value of 41.17 dB with an SSIM of 0.9993. Conversely, the RDHEI algorithm, when recovering the same watermark capacity, yields a PSNR value of 28.59 dB and an SSIM of 0.9435.

Discussion

Through in-depth experimental data analysis, the proposed BCC-BPR-ICWE algorithm exhibits a high watermark capacity and very low file incremental characteristics, enhancing the efficiency and convenience of the watermark embedding process. The algorithm ensures the integrity and security of copyright information while preserving the quality of the work without loss. Additionally, the image structure post-bit plane rearrangement is more concise and distinctive, facilitating subsequent image processing and analysis.

Furthermore, the algorithm efficiently recovers images while retaining the original tracking features, even after watermark embedding and subsequent attacks. The watermark embedding technique proves effective in protecting and authenticating digital artwork, offering artists and copyright owners a means to incorporate copyright information into their works, thereby preventing illegal copying and unauthorized distribution. This not only safeguards the legitimate rights and interests of artists but also promotes the healthy and orderly development of the art market.

Importantly, from the perspective of copyright protection, the algorithm in this article enables full-pixel watermark embedding, significantly enhancing the effectiveness of artwork copyright protection. With each pixel carrying copyright information, illegal copying or tampering of artworks becomes more easily recognized and traced. This approach not only combats piracy effectively but also provides comprehensive and reliable support for artists and copyright owners in defending their rights.

It’s worth noting that full-pixel watermark embedding may impact the storage and transmission performance of artwork resource management systems. As each pixel carries additional watermark information, the system’s data processing requirements increase significantly. However, the application of bit-plane rearrangement compresses the image, effectively eliminating redundant information and reducing storage space and transmission time.