A video screen projection protection method and device

Abstract

The present invention discloses a method and device for protecting video projection, wherein the method includes constructing a feature area for the projected video image, wherein the feature scale is determined when constructing the feature area, and the position, scale and direction of the feature point are determined based on the feature scale; a watermark is added to the video image in the feature area based on a pre-established watermark embedding network to obtain a watermark image; and an attack simulation is performed based on the watermark image. By constructing a feature area that is not only the feature point coordinates, and adding watermarks and attack simulations based on a network model, watermark embedding of the video projection is achieved, laying the foundation for subsequent leak tracing.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a method and device for protecting video projection. Background Art

[0002] The current common method for preventing video projection from being filmed is to use devices or technologies that can interfere with filming. These devices or technologies are usually called anti-surveillance devices or anti-surveillance technologies. For example, reflective materials (reflect light and cause the camera image to be blurred or flickering), infrared transmitters (transmit infrared light that interferes with infrared cameras, making it impossible for the camera to capture images normally), camera jammers (transmit electromagnetic signals or light signals of a specific frequency to interfere with the camera's work), etc. Although this method can effectively prevent candid photography, it cannot achieve the traceability function.

[0003] Another method is to use anti-peeping software. Anti-peeping screen software is a software solution used to protect the information on the computer screen from being viewed by others. This type of software aims to reduce the risk of screen content being leaked through biometric technology. It is especially suitable for public places or office environments to help protect users' privacy and sensitive information. For example, EyeGuardian is an anti-peeping screen software based on Windows Hello. It can detect when the user leaves the computer and automatically locks the screen to protect user privacy. However, anti-peeping screen software only provides a software-level supplement, which increases information security in an intelligent way. It cannot completely replace physical anti-peeping screen films or privacy filters, nor can it realize traceability functions.

[0004] Another method is dynamic encryption and DRM (digital rights management) technology, which combines dynamic encryption and DRM systems to ensure that only authorized devices and users can access and play content. However, this method is not very universal and is not suitable for public occasions. Summary of the invention

[0005] The present application provides a method and device for protecting video projection to solve the problems existing in the related art.

[0006] In a first aspect, the present invention provides a video projection method, comprising constructing a feature area for a projected video image, wherein a feature scale is determined when constructing the feature area, and a position, scale and direction of a feature point are determined based on the feature scale; a watermark is added to the video image in the feature area based on a pre-established watermark embedding network to obtain a watermark image; and an attack simulation is performed based on the watermark image.

[0007] Optionally, constructing a feature region for the projected video image includes: performing Gaussian filtering on the projected image based on a set Gaussian kernel and a width of the Gaussian function, and performing convolution processing after the Gaussian filtering. ; Based on a predefined second-order matrix of adaptive scale, the convolution processed image is processed to calculate the Gaussian multi-scale spatial representation at the input scale, wherein the second-order matrix is ,in, The standard deviation is Gaussian function, , and Respectively expressed in scale The derivatives of the Gaussian function in each direction under the Gaussian multi-scale space representation are obtained; a final feature scale is determined based on the Gaussian multi-scale space representation; and a position, scale and direction of the feature point are determined based on the final feature scale.

[0008] Optionally, the adding of watermarks to the video image in the feature area based on the pre-established watermark embedding network comprises: processing the video image based on a preset convolution kernel; connecting the processing result to a watermark message matrix, wherein the watermark message matrix is ​​reshaped by a binary watermark message string; processing the connection result by a preset number of residual blocks, and outputting a vector containing watermark information after the processing by the preset number of residual blocks; based on Perform inverse discrete cosine transform on the output vector to obtain parameter watermark information, where: is an intensity factor used to adjust the strength of the watermark information and the image quality. is a vector containing watermark information; the residual watermark information is superimposed on the video image to obtain the final watermark image.

[0009] Optionally, performing attack simulation based on the watermark image includes simulating an attack on a digital domain:

[0010] Randomly add noise attacks to the watermark image to simulate noise attacks; generate cropped random squares on the watermark image to simulate the situation where the watermark image is irregularly truncated; simulate compression attacks based on differentiable approximation methods.

[0011] Optionally, performing attack simulation based on the watermark image includes performing simulation based on a trained unsupervised neural network, wherein dense feature blocks in the unsupervised neural network have different scales to simulate attacks at different scales.

[0012] Optionally, the method further comprises minimizing the gap between the watermarked image and the distorted image caused by the attack in a manner based on updating parameters of the enhanced network.

[0013] Optionally, the method further includes tracing the watermark extraction of the watermark image: extracting the watermark embedded in the watermark image based on a fully convolutional network, wherein when the extraction network is trained for watermark extraction, minimizing the difference between the input watermark and the output watermark is used as the extraction target to update the parameters.

[0014] In a second aspect, the present invention provides a protection device for video projection, comprising a feature area construction unit, configured to construct a feature area for the projected video image; a watermark adding unit, configured to add a watermark to the video image in the feature area based on a pre-established watermark embedding network to obtain a watermark image; and an attack simulation unit, configured to perform attack simulation based on the watermark image.

[0015] In a third aspect, the present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the method described in any one of the first aspects is implemented.

[0016] In a fourth aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method provided in the first aspect when executing the program.

[0017] The present invention discloses a method and device for protecting video projection, wherein the method includes constructing a feature area for the projected video image, wherein the feature scale is determined when constructing the feature area, and the position, scale and direction of the feature point are determined based on the feature scale; a watermark is added to the video image in the feature area based on a pre-established watermark embedding network to obtain a watermark image; and an attack simulation is performed based on the watermark image. By constructing a feature area that is not only the feature point coordinates, and adding watermarks and attack simulations based on a network model. Watermark embedding of video projection is achieved, which further prevents leakage and lays the foundation for subsequent leakage tracing. The shortcomings existing in the relevant technology are overcome. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0019] Figure 1 A video screen projection method is provided for this application;

[0020] Figure 2 A method corresponding to the present application Figure 1 Schematic diagram of an electronic device. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0022] The following is combined with Figure 1 The following is an exemplary description of a method for protecting video projection. The method includes the following steps:

[0023] Step 101: Construct a feature area for the projected video image.

[0024] In this embodiment, the embedding of watermark information is the core part of digital watermarking technology, and constructing feature regions is a key step in this process. Feature points play an important role in digital watermarking technology. They are regarded as markers of the watermark embedding position, which helps to accurately locate the embedding position of the watermark information. However, relying solely on the coordinate information of feature points is not enough to deal with the scaling problems caused by attacks such as video projection.

[0025] As an optional implementation method of this embodiment, constructing a feature area for the projected video image includes: performing convolution processing on the projected image based on Gaussian filtering; processing the convolved image based on a predefined second-order matrix of adaptive scale to calculate the Gaussian multi-scale space representation at the input scale; determining the final feature scale based on the Gaussian multi-scale space representation; determining the position of the feature point based on the feature scale; and determining the position and direction of the feature point based on the final feature scale.

[0026] In this optional implementation, during the feature point extraction process, noise may affect its stability. To avoid this, Gaussian filtering is used to preprocess the brightness video image of the original data, which can not only reduce the impact of noise, but also help simulate low-pass filtering attacks, thereby filtering out some unstable feature points.

[0027] The Gaussian function is a probability density function that is represented as a bell-shaped curve in the spatial domain and has a Gaussian distribution. A two-dimensional Gaussian function It can be obtained by multiplying two one-dimensional Gaussian functions, and its mathematical expression is as follows:

[0028]

[0029] in, is the standard deviation, which determines the width of the Gaussian function. In this embodiment, in order to achieve the desired effect, a Gaussian kernel of size 6×6 can be selected. Set to 2. This parameter is based on experimental verification results and can achieve the best preprocessing effect.

[0030] Image convolution is a basic operation in digital image processing, which is used to implement various transformations and filtering of images. The formula for convolution operation is as follows:

[0031] Furthermore, in order to achieve scale invariance of video projection, a scale space representation of Harris detection strength is established. Harris corner detection is based on a specific image descriptor, namely the second-order moment matrix. This matrix can reflect the gradient changes in the local area of ​​the image. In order to make the matrix independent of the image scale, a scale-adaptive second-order moment matrix is ​​defined:

[0032] ,in, The standard deviation is The Gaussian function of , and Respectively expressed in scale The derivative of the Gaussian function in all directions.

[0033] Furthermore, to calculate the input scale The Gaussian multi-scale space representation under is expressed using the following formula: ,in, is the preprocessed image.

[0034] In this way, an effective Harris detection intensity scale space representation can be established at a predetermined scale, thereby achieving scale invariance. This method not only improves the robustness of feature point detection, but also provides a more accurate and reliable basis for subsequent feature matching and video image recognition.

[0035] Furthermore, through automatic feature scale selection, feature information of different scales can be processed more flexibly, thereby improving the robustness and accuracy of feature extraction. It is a method for automatically determining feature scales based on a given function. By calculating the results of the Laplacian-of-Gaussians (LoG) operator at each scale and finding the extreme point, the scale corresponding to the extreme point is used as the feature scale, and these candidate values ​​are further screened to determine the final feature scale. , The standard deviation is The Gaussian function of and denote the second-order partial derivatives in the x and y directions, respectively.

[0036] Furthermore, an iterative method is used to determine the position and scale of the feature points. After all points are calculated, the new point set is the Harris-Laplace feature point, which can be robust to the distortion and scale change of the projected image quality. In order to ensure the stability of the feature points to the rotation of the projected image, the SURF feature direction algorithm is used to assign a specific direction to each feature point. This process is crucial for scale, rotation and illumination invariance in image processing and feature recognition.

[0037] In order to determine the direction of the feature point, it is first necessary to calculate the image characteristics around it. This includes taking the feature point as the center, determining a circular area with a radius of n times the feature scale, performing a Haar wavelet response operation on the image within the circular area, and using a weighting function to adjust the response values ​​at different positions during the operation. The weighting function can be a Gaussian function, and the standard deviation can be set to k times the feature scale. Exemplarily, a circular area with a radius of 6 times the feature scale is selected. In this area, a Haar wavelet response operation is performed on the image, and a weighting function is used to adjust the response values ​​at different positions. This weighting function is a Gaussian function, and its standard deviation is twice the feature scale. This weighting method helps to emphasize the image information near the feature point, thereby more accurately determining its direction.

[0038] In order to determine the dominant direction in the area, a rotating window method is used. Specifically, a fan-shaped window with a preset angle, such as π / 3, is set with the feature point as the center. Then, the window is rotated with a preset step size, such as 0.2 radians. At each angle, the horizontal and vertical responses of the Haar wavelet in the window are added to obtain a vector. This vector not only represents the image characteristics in the area, but also provides direction information. In order to calculate the direction, the inverse tangent function is used. This function determines the direction based on the ratio of the horizontal response to the vertical response. In this way, the specific direction of each feature point can be obtained:

[0039]

[0040] in, represents the sum of the response values, Represents the corresponding direction.

[0041] Step 102: adding a watermark to the video image in the feature area based on a pre-established watermark embedding network to obtain a watermark image.

[0042] In this embodiment, watermark extraction can be performed based on Link-Net.

[0043] As an optional implementation method of this embodiment, adding a watermark to the video image in the feature area based on a pre-established watermark embedding network includes: processing the video image based on a preset convolution kernel; connecting the processing result to a watermark message matrix, wherein the watermark message matrix is ​​reshaped by a binary watermark message character string; processing the connection result by a preset number of residual blocks, and outputting a vector containing watermark information after processing by a preset number of residual blocks; performing an inverse discrete cosine transform on the output vector to obtain parameter watermark information; and superimposing the residual watermark information with the video image to obtain a final watermark image.

[0044] In this optional implementation, a 6×6 convolution kernel can be used for processing. Subsequently, the binary watermark message string is reshaped into a watermark message matrix and concatenated with the result of the first step. This concatenated result will pass through several residual blocks, each of which consists of a fully convolutional layer and a maximum pooling layer. The main function of these residual blocks is to extract and fuse different feature maps to provide rich feature information for subsequent watermark embedding. After a series of residual block processing, the network outputs a tensor containing the watermark information: In this network, the input and output of each residual block are closely connected. This design helps the network to make full use of different feature information to fuse watermark information. In this way, the high redundancy of the watermark can be ensured even in the face of some distortion. At the same time, due to the small number of network parameters, the training efficiency is significantly improved during training.

[0045] Furthermore, the output Vq of the embedded network is processed by inverse discrete cosine transform (IDCT) to obtain the residual watermark information Mw: ,in, is a strength factor used to adjust the strength of the watermark information and the image quality. The value of can balance the robustness of the watermark and the perceptual quality of the image.

[0046] Furthermore, the extracted watermark information With the original spanning image Superimpose to get the final watermark image : .

[0047] Step 103: Perform attack simulation based on the watermark image.

[0048] In this embodiment, in order to improve the robustness and invisibility of the screen projection watermark, a noise layer is introduced to simulate the real screen capture process. Although most single distortions can be easily simulated in the network, the actual photography transmission or light field information transmission involves more complex features. In the screen capture process, the watermark image may be affected by various complex patterns, which is significantly different from a simple attack combination.

[0049] As an optional implementation method of this embodiment, attack simulation based on the watermark image includes simulating attacks on the digital domain: randomly adding noise attacks to the watermark image to simulate noise attacks; generating cropped random squares on the watermark image to simulate the situation where the watermark image is irregularly truncated; simulating compression attacks based on a differentiable approximation method.

[0050] In this optional implementation, in order to more accurately simulate this series of distortions, they are divided into the following two main categories. The first type of distortion is a common attack in the digital domain, similar to a common noise attack. Since they are relatively easy to reproduce, they provide a basis for simulation. Specifically, the following three attacks are included: Noise attack: white noise is randomly added to the watermark image. In addition, Gaussian noise and salt and pepper noise are also randomly applied to image pixels to further increase the diversity of noise. Cropping attack: Generate cropped random squares on the watermark image to simulate the situation where the image is irregularly truncated. JPEG compression: Use a differentiable approximation method to simulate JPEG compression attacks. This method ensures that the distortion in the compression process can be accurately captured and simulated. By combining these attacks, the noise layer can effectively simulate common distortion conditions in the digital domain, thereby providing a more realistic input for subsequent processing.

[0051] As an optional implementation of this embodiment, performing attack simulation based on the watermark image includes performing simulation based on a trained unsupervised neural network, wherein the dense feature blocks in the unsupervised neural network have different scales to simulate attacks at different scales.

[0052] In this optional implementation, unlike the aforementioned distortion, the second type of distortion has a significant impact on screen capture watermark extraction. A key difference is that the image displayed on the screen is not a point-to-point pixel in the digital domain. Instead, each pixel is a light source propagating in free space. This characteristic makes the display emission function, the spectral sensitivity of the camera, and the camera's viewing angle important factors that affect the pixels captured by the camera. In order to accurately simulate this process, the concept of the camera display transfer function (CDTF) is adopted. CDTF is a complex function that can accurately describe the relationship between pixels and actual lighting based on photometric and radiometric effects.

[0053] To approximate this complex transfer process, a trained unsupervised neural network is utilized, which is characterized by its dense feature blocks of different scales, which enables the network to better capture and simulate illumination changes and distortions at different scales.

[0054] Due to differences in lighting conditions, camera positioning, focus range and other factors in the actual environment, deviations are inevitable. w The attack eventually leads to the distorted image I d By introducing a noise layer and training and adjusting it appropriately, the robustness of the network can be improved, making it better adaptable to the challenges and changes in various practical application scenarios.

[0055] As an optional implementation of this embodiment, the method further includes minimizing the gap between the watermark image and the distorted image caused by the attack in a manner of updating parameters based on an enhanced network.

[0056] In this optional implementation, the enhancement network is designed to mitigate the effects of optical interference and enhance the information of the original cross-screen image. Optical interference is the most important threat in the screen photography transmission process, and the cause of this interference mainly comes from frequency interference. In current work, the impact on the image is usually modeled as follows: ,in, Representation scale and frequency The image texture component.

[0057] The distorted image I d Downsampled to three different scales, the feature maps of each scale are passed through dense blocks for feature extraction. The dense blocks contain multiple 3×3 dilated convolutional layers activated by ReLU. Dilated convolutions are used to further increase the receptive field and implicitly learn the spectrum in the DCT domain. The IDCT layer is basically the same as the IDCT layer in the encoder. Then, they all pass through the global average pooling layer to extract global features. The full convolution block is used to learn and restore the global color. After outputting the results of each branch, they are upsampled and connected to the previous branch in sequence. The same structure is executed to obtain more detailed results and eliminate the effects caused by the connection.

[0058] In addition, the spatial transformer network (STN) is a network structure used to enhance the robustness of the model to spatial transformations (such as translation, rotation, and scaling). In the initial block of the network, the spatial transformation parameters of the input image are learned to appropriately transform the input image, thereby reducing the interference caused by pixel displacement in subsequent processing steps, which is used to combat small pixel displacements. In order to reduce the interference of screen capture distortion and improve the accuracy of extracting watermark information, the goal of the enhancement network is to minimize the watermark image I by updating the parameters. wWith enhanced network output I en The gap between EN : .

[0059] In summary, STN first performs spatial correction on the input image to reduce the impact of geometric distortion on subsequent processing. The noise layer then simulates actual noise attacks to ensure that the watermark algorithm can remain robust in real environments. The enhancement network finally restores the features of the watermark image to improve the accuracy and reliability of watermark extraction.

[0060] As an optional implementation of this embodiment, the method also includes tracing the watermark extraction of the watermark image: extracting the watermark embedded in the watermark image based on a fully convolutional network, wherein when the extraction network is trained for watermark extraction, minimizing the difference between the input watermark and the output watermark is used as the extraction target to update the parameters.

[0061] In this optional implementation, the fully convolutional network (FCN) is used to extract the watermark embedded in the projection watermark image. The backbone of the extraction network is the same as the embedding network. It is worth noting that the last block of the extraction network uses Sigmoid as the activation function to normalize the data to the range of [0, 1] to output the message matrix. The goal of the extraction network is to update the parameters To minimize the difference L between Mo and Mi EX : .

[0062] In this embodiment, the embedding network based on the Link-Net structure realizes the encoding process of the video projection image and the watermark information. The extraction network based on the FCN structure recovers the watermark information from the watermark image. The enhancement network based on the U-Net structure enhances the features of the distorted image. The noise layer simulates noise attacks in various real scenes to ensure that the watermark information is not easily tampered with or deleted.

[0063] While the watermark extraction accuracy is high, it retains good watermark image quality and is not easy to be destroyed or tampered with. It is based on various types of large screens, including splicing screens of various specifications and large-size LCD TV scenes. It has the ability to trace the source and collect evidence across optical channels, which has a strong deterrent effect on illegal cross-screen shooting and leaking of information, and provides technical support for subsequent evidence collection and accountability.

[0064] The present application also provides a protection device for video projection, including a feature area construction unit, configured to construct a feature area for the projected video image; a watermark adding unit, configured to add a watermark to the video image in the feature area based on a pre-established watermark embedding network to obtain a watermark image; and an attack simulation unit, configured to perform attack simulation based on the watermark image.

[0065] As an optional implementation of this embodiment, constructing a feature area for the projected video image includes: performing Gaussian filtering on the projected image based on a set Gaussian kernel and a width of the Gaussian function, and performing convolution processing after Gaussian filtering. ; Based on a predefined second-order matrix of adaptive scale, the convolution processed image is processed to calculate the Gaussian multi-scale spatial representation at the input scale, wherein the second-order matrix is ,in, The standard deviation is Gaussian function, , and Respectively expressed in scale The derivatives of the Gaussian function in each direction under the Gaussian multi-scale space representation are obtained; a final feature scale is determined based on the Gaussian multi-scale space representation; and a position, scale and direction of the feature point are determined based on the final feature scale.

[0066] As an optional implementation of this embodiment, the adding of watermarks to the video image in the feature area based on the pre-established watermark embedding network includes: processing the video image based on a preset convolution kernel; connecting the processing result to a watermark message matrix, wherein the watermark message matrix is ​​reshaped by a binary watermark message string; processing the connection result by a preset number of residual blocks, and outputting a vector containing watermark information after the preset number of residual blocks are processed; based on Perform inverse discrete cosine transform on the output vector to obtain parameter watermark information, where: is an intensity factor used to adjust the strength of the watermark information and the image quality. is a vector containing watermark information; the residual watermark information is superimposed on the video image to obtain the final watermark image.

[0067] As an optional implementation method of this embodiment, attack simulation based on the watermark image includes simulating attacks on the digital domain: randomly adding noise attacks to the watermark image to simulate noise attacks; generating cropped random squares on the watermark image to simulate the situation where the watermark image is irregularly truncated; simulating compression attacks based on a differentiable approximation method.

[0068] As an optional implementation of this embodiment, performing attack simulation based on the watermark image includes performing simulation based on a trained unsupervised neural network, wherein the dense feature blocks in the unsupervised neural network have different scales to simulate attacks at different scales.

[0069] As an optional implementation of this embodiment, the method further includes minimizing the gap between the watermark image and the distorted image caused by the attack in a manner of updating parameters based on an enhanced network.

[0070] As an optional implementation of this embodiment, the method also includes tracing the watermark extraction of the watermark image: extracting the watermark embedded in the watermark image based on a fully convolutional network, wherein when the extraction network is trained for watermark extraction, minimizing the difference between the input watermark and the output watermark is used as the extraction target to update the parameters.

[0071] Figure 2 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as Figure 2 As shown, the electronic device 50 includes: a processor 501 (processor), a memory 502 (memory) and a bus 503;

[0072] The processor 501 and the memory 502 communicate with each other via the bus 503 ; the processor 501 is used to call program instructions in the memory 502 to execute the methods provided by the above-mentioned method embodiments.

[0073] This embodiment provides a non-transitory computer-readable storage medium, which stores computer instructions. The computer instructions enable a computer to execute the methods provided by the above method embodiments.

[0074] A person skilled in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various storage media that can store program codes.

[0075] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0076] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0077] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for protecting video projection, characterized in that: include: Constructing a feature area for the projected video image, wherein a feature scale is determined when constructing the feature area, and a position, scale, and direction of a feature point are determined based on the feature scale; Adding a watermark to the video image in the feature area based on a pre-established watermark embedding network to obtain a watermarked image; Performing attack simulation based on the watermark image; The feature areas constructed for the projected video image include: Based on the set Gaussian kernel and the width of the Gaussian function, the image to be projected is Gaussian filtered, and convolution is performed after Gaussian filtering L′0(x, y)=H G ×L0(x,y), where H G represents the Gaussian kernel, L0(x,y) represents the original image, L′0(x,y) represents generating a new image, (x,y) means that each pixel has a corresponding coordinate position; The convolution-processed image is processed based on a predefined second-order matrix of adaptive scale to calculate the Gaussian multi-scale spatial representation at the input scale, wherein the second-order matrix is Among them, G(σ I ) is the standard deviation of σ I Gaussian function, L x L y (x,y,σ D )and Respectively represent the derivatives of the Gaussian function in each direction at scale L; Determining a final feature scale based on the Gaussian multi-scale spatial representation; The position, scale and orientation of the feature points are determined based on the final feature scale.

2. The method for protecting video projection according to claim 1, characterized in that: Adding a watermark to the video image in the feature area based on a pre-established watermark embedding network comprises: Process the video image based on the preset convolution kernel; Connecting the result of the processing to a watermark message matrix, wherein the watermark message matrix is ​​reshaped from a binary watermark message string; The result of the connection is processed by a preset number of residual blocks, and a vector containing watermark information is output after being processed by the preset number of residual blocks; Based on M w =α×IDCT(V q ) performs an inverse discrete cosine transform on the output vector to obtain parameter watermark information, where α is an intensity factor used to adjust the intensity of the watermark information and the image quality, V q is a vector containing watermark information; The residual watermark information is superimposed on the video image to obtain the final watermark image.

3. The video screen projection protection method according to claim 1, characterized in that: The attack simulation based on the watermark image includes simulating the attack on the digital domain: Randomly add noise attack to the watermark image to simulate noise attack; Generate cropped random squares on the watermark image to simulate the situation where the watermark image is irregularly truncated; Simulating compression attacks based on differentiable approximation methods.

4. The method for protecting video projection according to claim 1, characterized in that: Performing attack simulation based on the watermark image includes performing simulation based on a trained unsupervised neural network, wherein the dense feature blocks in the unsupervised neural network have different scales to simulate attacks at different scales.

5. The method for protecting video projection according to claim 3 or 4, characterized in that: The method also includes minimizing the gap between the watermarked image and the distorted image caused by the attack in a manner based on updating parameters of the enhancement network.

6. The method for protecting video projection according to claim 1, characterized in that: The method further comprises tracing the watermark extraction from the watermark image: A watermark embedded in a watermark image is extracted based on a fully convolutional network, wherein the extraction network, when being trained for watermark extraction, updates parameters by minimizing the difference between the input watermark and the output watermark as the extraction target.

7. A video screen projection protection device, characterized in that: include A feature region construction unit is configured to construct a feature region for the projected video image; A watermark adding unit is configured to add a watermark to the video image in the feature area based on a pre-established watermark embedding network to obtain a watermark image; an attack simulation unit, configured to perform attack simulation based on the watermark image; Among them, constructing feature areas for the projected video image includes: Based on the set Gaussian kernel and the width of the Gaussian function, the image to be projected is Gaussian filtered, and convolution is performed after Gaussian filtering L′0(x, y)=H G ×L0(x,y), where H G represents the Gaussian kernel, L0(x,y) represents the original image, L′0(x,y) represents generating a new image, (x,y) means that each pixel has a corresponding coordinate position; The convolution-processed image is processed based on a predefined second-order matrix of adaptive scale to calculate the Gaussian multi-scale spatial representation at the input scale, wherein the second-order matrix is Among them, G(σ I ) is the standard deviation of σ I Gaussian function, L x L y (x,y,σ D )and Respectively represent the derivatives of the Gaussian function in each direction at scale L; Determining a final feature scale based on the Gaussian multi-scale spatial representation; The position, scale and orientation of the feature points are determined based on the final feature scale.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

Images