Chart image depth watermarking method based on serial embedding and parallel extraction

Abstract

The invention discloses a chart image deep watermarking method based on serial embedding and parallel extraction, which comprises the following steps of: firstly, carrying out first-stage serial training on a watermark model, and training and finely adjusting the watermark model by using each noise layer one by one to obtain a watermark encoder capable of embedding a watermark feature with various distortion resistance into a chart image; and then parallel training of a second stage is carried out, the watermark encoders obtained in the serial training stage are fixed, a plurality of watermark decoders are simultaneously trained in parallel, and each watermark encoder carries out watermark information extraction for one specific distortion. According to the method, while the watermark embedding is ensured not to influence the quality of the chart image, the copyright protection and leakage prevention functions of the chart image can be realized, so that the intellectual property and security protection of the chart image can be enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision, and in particular to a chart image deep watermarking method based on serial embedding and parallel extraction. Background Art

[0002] Various chart images, such as line charts, pie charts, bar charts, etc., can provide a visual display of data, allowing people to understand and analyze data more intuitively, thus playing a vital role in various important scenarios such as financial analysis and business management. Due to the high value of chart images, people urgently need to find effective ways to protect the legal copyright of chart images and prevent unauthorized leakage of chart images. Image watermarking technology can effectively prevent copyright infringement and leaks, and has always received widespread attention from academia and industry. In recent years, with the development of deep learning technology, many image watermarking methods based on deep learning have been continuously proposed, and have shown significantly better performance than previous traditional image watermarking methods. The previous deep watermarking scheme framework was mainly designed for natural images, which mainly includes watermark encoders, noise layers, and watermark decoders. The watermark encoder needs to embed the watermark message into the image in a way that is invisible to the human eye, and try to ensure that the visual quality of the image is not reduced by the embedding of the watermark; the noise layer usually appears between the watermark encoder and the decoder. It simulates the distortion that may occur in the image during channel transmission and adds noise to the watermarked image, thereby adding interference to the training process to help the model improve its robustness to related distortions through training. By using various noise layers, the model can flexibly obtain various corresponding robustness; finally, the watermark decoder needs to process the received watermarked image that may have been interfered with and accurately extract the watermark message therein.

[0003] Following the design framework of the above-mentioned deep learning watermarking method for natural images, some deep learning watermarking schemes for graphic images have also been proposed. However, such frameworks still have some limitations. Such schemes rely on a combined noise layer containing multiple distortions to train the watermark model to improve the model's multiple robustness to multiple image distortions. Each sample batch in the training will be randomly interfered with by a certain type of distortion contained in the combined noise layer. Such variability will pose a challenge to the convergence of watermark model training, making it difficult for the model to learn how to embed watermark features suitable for various distortions into the image. Therefore, previous deep learning watermarking schemes for graphic images designed based on this framework still have room for improvement in obtaining multiple robustness to multiple distortions.

[0004] At the same time, as a special type of image, chart images may encounter special distortions in practical applications, which are not considered by previous chart image watermarking schemes. In addition, for various distortions that may occur during channel transmission and are not considered during training, effective countermeasures need to be designed to avoid the serious threat posed by these distortions that were not considered during training to the accurate extraction of watermarks. This is also an important aspect that the existing deep learning chart image watermarking scheme needs to improve. Summary of the invention

[0005] The purpose of the present invention is to provide a deep watermarking method for chart images based on serial embedding and parallel extraction, which can ensure that watermark embedding does not affect the quality of the chart image, while also realizing the copyright protection and leakage prevention functions of the chart image, thereby helping to strengthen the intellectual property and security protection of the chart image.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A chart image deep watermarking method based on serial embedding and parallel extraction, the method comprising:

[0008] Step 1, perform the first phase serial training on the watermark model, use each noise layer to train and fine-tune the watermark model one by one, and obtain a watermark encoder that can embed watermark features with the ability to resist multiple distortions into the chart image;

[0009] Step 2: Perform the second stage of parallel training, fix the watermark encoder obtained in the serial training stage, and train multiple watermark decoders in parallel. Each watermark encoder extracts watermark information for a specific distortion.

[0010] It can be seen from the technical solution provided by the above-mentioned invention that the above-mentioned method can ensure that watermark embedding will not affect the quality of the chart image while realizing the copyright protection and leakage prevention functions of the chart image, thereby helping to strengthen the intellectual property and security protection of the chart image, and has strong practicality in real scenarios where it is necessary to protect the copyright of the chart image and prevent the leakage of the chart image. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0012] Figure 1 A schematic flow chart of a method for deep watermarking of a chart image based on serial embedding and parallel extraction provided by an embodiment of the present invention;

[0013] Figure 2 A schematic diagram of a noise layer based on a combination of unknown distortion simulations according to an embodiment of the present invention;

[0014] Figure 3 Schematic diagram of two special noise layers for graphic images according to an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments, which does not constitute a limitation of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] like Figure 1 FIG. 1 is a flow chart of a method for deep watermarking of a chart image based on serial embedding and parallel extraction provided by an embodiment of the present invention. The method comprises:

[0017] Step 1, perform the first phase serial training on the watermark model, use each noise layer to train and fine-tune the watermark model one by one, and obtain a watermark encoder that can embed watermark features with the ability to resist multiple distortions into the chart image;

[0018] In this step, in the serial training phase, the watermark encoder, watermark decoder and discriminator are first initialized; then a distortion is randomly selected from a total of n distortions and set as the first noise layer used to participate in end-to-end training, and it is used to train the initialized watermark model, including the watermark encoder and watermark decoder;

[0019] Input carrier chart image I to the watermark encoder o and watermark information M, the watermark encoder outputs the watermarked chart image I en ; Then the watermarked chart image I en will be disturbed by the first noise layer and produce a distorted watermarked chart image I no ;

[0020] Input I to the watermark decoder no , the watermark information M′ is extracted by the watermark decoder output. In this process, the discriminator A is used to try to distinguish I o with I en , when the discriminator A cannot distinguish the difference between the two, it means that the watermarked chart image I en The visual quality of the watermark is up to standard; the next n-1 distortions will be used serially one by one to continuously fine-tune the watermark model;

[0021] The loss function used to train and fine-tune the watermark model in the serial training phase consists of three parts. The first part is the image quality loss L guided by the JND mask. Enc , by the carrier chart image I o Apply the JND algorithm to get the corresponding JND mask, which is used to represent the sensitivity of the human visual system to different areas in the chart image; the image quality loss L Enc The calculation method is:

[0022] L Enc =MSE(I en ·mask, I o ·mask)

[0023] Where MSE() stands for minimum mean square error loss; mask represents the carrier chart image I o The corresponding JND mask, the smaller the value of a position, the lower the sensitivity of the human eye to the corresponding position in the original chart image; therefore, in the training and fine-tuning process, in order to Enc To minimize, the watermark encoder will tend to modify the areas that are less sensitive to the human eye under the guidance of the JND mask, thereby improving the visual quality of the watermarked chart image;

[0024] The second part is L Dec , which is used to measure the gap between the original watermark information and the extracted watermark information. Dec It can improve robustness and ensure accurate extraction of watermark information. The loss function L Dec The calculation method is:

[0025] L Dec =MSE(M,M′)

[0026] Where M is the original watermark information; M′ is the extracted watermark information;

[0027] In the third part, the discriminator A tries to identify the watermarked image I en Perform binary classification to distinguish I en Is it a watermarked chart image or an original chart image, the corresponding loss function L D The calculation method is:

[0028] L D =BcE(1,A(I en ))

[0029] Where BCE() represents the binary cross entropy loss; A(I en ) represents the discriminator A for the watermarked chart image I enAs the output of the input, if the discriminator A thinks that the input image is an image without watermark, the output value is 1, if the discriminator A thinks that the input image is an image with watermark, the output value is 0; therefore, in order to minimize L during training and fine-tuning D , the watermark encoder will be trained to minimize the difference between the watermarked chart image and the original chart image, thereby improving the visual quality of the watermarked image;

[0030] Therefore, the total loss function L is calculated as:

[0031] L=α1L D +α2L Enc +α3L Dec

[0032] Where α1, α2 and α3 represent L D , L Enc and L Dec The respective weights in the calculation of L, in this example they are set to 0.001, 1.0 and 3.0 respectively;

[0033] At the same time, in order to obtain better training results, the discriminator A will also be continuously trained and optimized in this process. The loss function of training the discriminator A is:

[0034] L A =BCE(0,A(I en ))+BCE(1,A(I o ))

[0035] Among them A(I o ) represents the discriminator A for the carrier graph image I o as the output when it is input; by minimizing L A , the discriminator A’s ability to discriminate whether a chart image contains a watermark will be improved, which will also help improve the training effect of the watermark model, thereby more effectively improving the ability of the watermark model through training and fine-tuning.

[0036] In the specific implementation, compared with natural images, chart images have special properties. The layout of chart images is relatively fixed (generally, they can be divided into two parts: data-related areas and background areas), and the texture is simpler. Only the data-related areas contain some areas with relatively complex textures suitable for watermark information embedding. Therefore, the areas suitable for watermark embedding in chart images lack variability, and are mainly concentrated in the data-related areas regardless of the distortion. Such characteristics make the watermark feature patterns required to deal with different distortions on chart images have a large overlap. Therefore, in the serial training process, the watermark feature patterns obtained by training to deal with the image distortions that have been used in a series of image distortions involved in the serial training can still be retained to a large extent when fine-tuning with the next image distortion. The watermark feature patterns required for subsequent fine-tuning can be based on the results obtained from previous training and fine-tuning, which is conducive to the watermark model to better converge to the ideal result that is robust to multiple distortions through serial training.

[0037] Step 2: Perform the second stage of parallel training, fix the watermark encoder obtained in the serial training stage, and train multiple watermark decoders in parallel. Each watermark encoder extracts watermark information for a specific distortion.

[0038] In this step, in the parallel training stage, the watermark encoder obtained by the training and fine-tuning in the first stage will be saved and fixed, and multiple watermark decoders matching it will be retrained in parallel. Each watermark decoder is trained under a specific distortion condition to ensure that there is a corresponding watermark decoder under each different distortion condition to extract the watermark information in a targeted manner, further improving the robustness of the present invention against various image distortions.

[0039] Among them, in the parallel training stage, the loss function used in training is the loss function L in the first stage D ec.

[0040] In the specific implementation, in the actual application scenario, the information that can be obtained in advance about the types of distortion that the watermarked image may encounter during the channel propagation process is extremely scarce. If the preset degree of possible image distortion can be further reduced, it will have further use value in the actual scenario. Therefore, the embodiment of the present invention also proposes a noise layer based on the combination of unknown distortion simulation, such as Figure 2 FIG. 1 is a schematic diagram of a noise layer based on a combination of unknown distortion simulations according to an embodiment of the present invention, which aims to improve the robustness of the trained watermark model to cope with JPEG compression and various unknown image distortions not considered during training, and includes three parts:

[0041] 1) A trained agnostic noise layer based on a convolutional neural network, which can impose universal interference on the watermarked image without distinguishing specific types;

[0042] 2)JPEG compression operation;

[0043] 3) Identity operation without any image distortion;

[0044] The training loss function L of the unknown noise layer C built based on the convolutional neural network is C for:

[0045]

[0046] Among them I en is the watermarked chart image output by the watermark encoder; I′ en is the watermarked chart image after being disturbed by the unknown noise layer C; M is the original watermark information; M′ is the watermark decoder from I′ en The watermark information extracted from ||I′ en -I en || 2 The effect of the unknown noise layer C on the watermarked chart image I is measured. en The interference degree of the watermark decoder; MSE() stands for the minimum mean square error loss; MSE(M′, M) measures the impact of the watermark extraction accuracy of the watermark decoder after being processed by the unknown noise layer C; and is the loss function L C The weights of the two losses involved in ; in this example they are set to 15.0 and 1.0 respectively;

[0047] The unknown noise layer C is combined with the JPEG compression operation and the Identity operation without any image distortion to form a combination-based unknown distortion simulation noise layer;

[0048] When in use, for each batch of training samples, one of the three components contained in the combined unknown distortion simulation noise layer is randomly selected, and corresponding processing is applied to the watermarked chart image corresponding to the training sample batch; (for example, if the unknown noise layer C is drawn for this batch of training samples, the unknown noise layer C is used to apply type-indiscriminate distortion interference to the watermarked chart image; if the Identity operation is drawn, no distortion is applied), and the watermarked chart image after interference based on the combined unknown distortion simulation noise layer is output.

[0049] Therefore, it can be seen that the unknown distortion simulation noise layer based on the combination requires almost no prior knowledge of the image propagation channel conditions, and can well help the watermark model adapt to various types of distortion that may occur in the process of channel propagation, and greatly improve the robustness of the watermark model to various types of distortion that have not been experienced during training.

[0050] In addition, as a special type of image, chart images may encounter some special distortions for chart images in their application scenarios. However, previous deep learning watermarking methods for chart images have not taken into account the special distortions of chart images, thereby limiting their corresponding robustness in this regard and their scope of application in actual scenarios. In the parallel training phase of step 2, the present invention designs two special noise layers for chart images according to special application scenarios of chart images, including adding titles to chart images and overlapping two chart images; wherein, adding titles to chart images is intended to add some explanations to the original chart images in order to better display and analyze data; overlapping two chart images is a multi-chart mixing operation, which can effectively mix chart images of different forms into one image, thereby better displaying more different attributes, which is helpful for people to analyze and compare data.

[0051] like Figure 3 The diagram shows two special noise layers for chart images according to an embodiment of the present invention. Adding a noise layer corresponding to a title to a chart image is achieved by replacing certain areas in the chart image with rectangular blocks of random size and color. Specifically, for each chart image, a vertex coordinate of a rectangle in the chart image and its height and width values ​​are randomly generated to randomly select a rectangular area; then, a pixel value corresponding to a specific color is randomly generated to randomly select a color for filling the rectangle.

[0052] For the noise layer corresponding to the two overlapping chart images, the implementation method is to add the two chart images with weights, and the calculation method is:

[0053] I no =αI1+(1-α)I2

[0054] Among them, I no is the image obtained by overlapping two chart images; I1 and I2 are two chart images to be overlapped; α and 1-α are the weights of the two chart images when weighted addition is performed.

[0055] The special noise layer will participate in the parallel training of the watermark decoder, thereby improving the ability to cope with special distortions of chart images and further expanding the application scope of the present invention in practical scenarios.

[0056] The performance of the chart image watermarking model after the two-stage serial-parallel training of step 1 and step 2 above will be significantly better than that of previous similar schemes. In order to prove the effectiveness of the present invention, a comprehensive comparison is made with existing deep watermarking schemes from two aspects: robustness and visual quality of watermarked images. The baseline methods include TSDL (A Novel Two-stage Separable Deep Learning Framework for Practical Blind Watermarking), MBRS (MBRS: Enhancing Robustness of DNN-based Watermarking by Mini-Batch of Real and Simulated JPEG Compression, a deep learning image watermarking scheme that improves robustness by alternating between real or simulated JPEG compression during model training), and ChartStamp (ChartStamp: Robust Chart Embedding for Real-World Applications, a robust chart image watermarking scheme for real-world applications). In the experiment, 2635 publicly released chart images collected from the Internet were selected for watermark model training, and another 200 publicly released chart images collected from the Internet were used for watermark model performance evaluation.

[0057] For the visual quality of the watermarked image, the indicator used is the PSNR (peak signal-to-noise ratio) value used to evaluate the difference between the watermarked image and the original image. The higher its value, the better the visual quality of the watermarked image. The results are shown in Table 1:

[0058] Table 1 Comparison of visual quality of watermarked chart images in different schemes

[0059]

[0060] From the average results calculated on 200 test images, the solution proposed in the embodiment of the present invention can achieve the highest PSNR value, which fully demonstrates the obvious advantages of the present invention over previous solutions in terms of the visual quality of watermarked chart images.

[0061] For robustness, the evaluation indicator used is the bit accuracy rate (BCR) of watermark extraction under various distortion conditions. The higher the BCR, the stronger the robustness when distortion occurs. The results are shown in Table 2:

[0062] Table 2 Comparison of watermark extraction bit accuracy of different schemes under various distortion conditions

[0063]

[0064] Judging from the average results calculated on 200 test images, for various general image distortions and two special distortions of chart images used in the test, the present invention has achieved significantly higher BCR values ​​than previous solutions, and can reach or approach 100.00% watermark extraction accuracy. These sufficient experimental results effectively illustrate the excellent robustness of the present invention in maintaining the most accurate watermark message extraction effect under various types of image distortion conditions.

[0065] It is worth noting that the contents not described in detail in the embodiments of the present invention belong to the prior art known to professional and technical personnel in the field.

[0066] In summary, the method described in the embodiment of the present invention has the following advantages:

[0067] 1. The present invention proposes a new chart image watermark model training process based on two-stage serial-parallel training, which is different from the previous chart image deep learning image watermarking scheme. By applying this two-stage training method, the watermark model trained by the present invention can achieve significantly better robustness and visual quality of watermarked images compared with previous similar schemes;

[0068] 2. The present invention proposes a noise layer based on unknown distortion simulation based on combination. By combining the use of an unknown noise layer based on a convolutional neural network, JPEG compression, and the situation where no image distortion occurs (Identity), it is almost unnecessary to know the propagation channel of the image in advance. It can help the watermark model adapt to various types of distortion that may occur in the process of channel propagation, and greatly improve the robustness of the watermark model to various types of distortion that have not been experienced during the training process.

[0069] 3. The present invention proposes two special noise layers for chart images, corresponding to the two special distortions of chart images: adding a title to a chart image and overlapping two chart images. These two noise layers are used in the training process of the watermark model to improve the corresponding robustness, further expanding the practical value of the present invention in the actual application scenarios of chart images.

[0070] In addition, a person skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be implemented by instructing related hardware through a program, and the corresponding program can be stored in a computer-readable storage medium. The above-mentioned storage medium can be a read-only memory, a disk or an optical disk, etc.

[0071] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims. The information disclosed in the background technology section of this article is only intended to deepen the understanding of the overall background technology of the present invention, and should not be regarded as an admission or in any form that the information constitutes prior art known to those skilled in the art.

Claims

1. A chart image deep watermarking method based on serial embedding and parallel extraction, characterized in that: The method comprises: Step 1, perform the first phase serial training on the watermark model, use each noise layer to train and fine-tune the watermark model one by one, and obtain a watermark encoder that can embed watermark features with the ability to resist multiple distortions into the chart image; Step 2: Perform the second stage of parallel training, fix the watermark encoder obtained in the serial training stage, and train multiple watermark decoders in parallel. Each watermark encoder extracts watermark information for a specific distortion.

2. According to claim 1, the method for deep watermarking of chart images based on serial embedding and parallel extraction is characterized in that: In step 1, in the serial training phase, the watermark encoder, watermark decoder and discriminator are first initialized; then a distortion is randomly selected from a total of n distortions and set as the first noise layer used to participate in end-to-end training, and it is used to train the initialized watermark model, including the watermark encoder and watermark decoder; Input carrier chart image I to the watermark encoder o and watermark information M, the watermark encoder outputs the watermarked chart image I en ; Then the watermarked chart image I en will be disturbed by the first noise layer and produce a distorted watermarked chart image I no ; Input I to the watermark decoder no , the watermark information M′ is extracted by the watermark decoder output. In this process, the discriminator A is used to try to distinguish I o with I en , when the discriminator A cannot distinguish the difference between the two, it means that the watermarked chart image I en The next n-1 distortions will be used serially one by one to continuously fine-tune the watermark model; The loss function used to train and fine-tune the watermark model in the serial training phase consists of three parts. The first part is the image quality loss L guided by the JND mask. Enc , by the carrier chart image I o Apply the JND algorithm to get the corresponding JND mask, which is used to represent the sensitivity of the human visual system to different areas in the chart image; the image quality loss L Enc The calculation method is: L Enc =MSE(I en ·mask,I o ·mask) Where MSE() stands for minimum mean square error loss; mask represents the carrier chart image I o In the corresponding JND mask, the smaller the value of a position is, the lower the sensitivity of the human eye to the corresponding position in the original chart image; The second part is L Dec , which is used to measure the gap between the original watermark information and the extracted watermark information. Dec It can improve robustness and ensure accurate extraction of watermark information. The loss function L Dec The calculation method is: L Dec =MSE(M,M′) Where M is the original watermark information; M is the extracted watermark information; In the third part, the discriminator A tries to identify the watermarked image I en Perform binary classification to distinguish I en Is it a watermarked chart image or an original chart image, the corresponding loss function L D The calculation method is: L D =BCE(1,A(I en )) Where BCE() represents the binary cross entropy loss; A(I en ) represents the discriminator A for the watermarked chart image I en As the output of the input, if the discriminator A believes that the input image is an image without a watermark, the output value is 1, if the discriminator A believes that the input image is an image with a watermark, the output value is 0; Therefore, the total loss function L is calculated as: L=α1L D +α2L Enc +α3L Dec Where α1, α2 and α3 represent L D , L Enc and L Dec The respective weights in the calculation process of L; At the same time, in order to obtain better training results, the discriminator A will also be continuously trained and optimized in this process. The loss function of training the discriminator A is: L A =BCE(0,A(I en ))+BCE(1,A(I o )) Among them A(I o ) represents the discriminator A for the carrier graph image I o as the output when it is input; by minimizing L A , the discriminator A’s ability to discriminate whether a graphic image contains a watermark will be improved.

3. The method for deep watermarking of chart images based on serial embedding and parallel extraction according to claim 2 is characterized in that: In step 2, in the parallel training phase, the watermark encoder trained and fine-tuned in the first phase will be saved and fixed, and multiple matching watermark decoders will be retrained in parallel. Each watermark decoder is trained under a specific distortion condition to ensure that there is a corresponding watermark decoder under each different distortion condition to extract the watermark information in a targeted manner. Among them, in the parallel training stage, the loss function used in training is the loss function L in the first stage Dec .

4. The method for deep watermarking of chart images based on serial embedding and parallel extraction according to claim 1 is characterized in that: In the serial training phase of step 1, a noise layer based on a combination of unknown distortion simulations is designed, which consists of three parts: 1) A trained agnostic noise layer based on a convolutional neural network, which can impose universal interference on the watermarked image without distinguishing specific types; 2)JPEG compression operation; 3) Identity operation without any image distortion; The training loss function L of the unknown noise layer C built based on the convolutional neural network is C for: Among them I en is the watermarked chart image output by the watermark encoder; I′ en is the watermarked chart image after being disturbed by the unknown noise layer C; M is the original watermark information; M′ is the watermark decoder from I′ en The watermark information extracted from ||I′ en -I en || 2 The effect of the unknown noise layer C on the watermarked chart image I is measured. en The degree of interference; MSE() stands for minimum mean square error loss; MSE(M′, M) measures the impact of the watermark extraction accuracy of the watermark decoder after being processed by the unknown noise layer C; and is the loss function L C The weights of the two parts of loss involved in; The unknown noise layer C is combined with the JPEG compression operation and the Identity operation without any image distortion to form a combination-based unknown distortion simulation noise layer; When in use, for each batch of training samples, one of the three components contained in the combined unknown distortion simulated noise layer is randomly selected, and corresponding processing is applied to the watermarked chart image corresponding to the training sample batch, and the watermarked chart image after interference based on the combined unknown distortion simulated noise layer is output.

5. The method for deep watermarking of chart images based on serial embedding and parallel extraction according to claim 1, characterized in that: In the parallel training phase of step 2, two special noise layers for chart images are designed for special application scenarios of chart images, including adding titles to chart images and overlapping two chart images, where: Adding a noise layer corresponding to the title to the chart image is achieved by replacing certain areas in the chart image with rectangular blocks of random size and color. Specifically, for each chart image, a vertex coordinate of a rectangle in the chart image and its height and width values ​​are randomly generated to randomly select a rectangular area; then, the color used to fill this rectangle is randomly selected by randomly generating pixel values ​​corresponding to a specific color; For the noise layer corresponding to the two overlapping chart images, the implementation method is to add the two chart images with weights, and the calculation method is: I no =αI1+(1-α)I2 Among them, I no is the image obtained by overlapping two chart images; I1 and I2 are two chart images to be overlapped; α and 1-α are the weights of the two chart images when weighted addition is performed.

Images