Method for obtaining moire pattern recognition image based on multiple modeling of multi-style image

Abstract

The invention discloses a method for acquiring a moire pattern recognition image based on multiple modeling of a multi-style image. The method comprises the following steps: S1, performing multiple modeling on an image; and S2, image recognition. According to the method, multiple and repeated training, verification and testing are carried out on the image based on multiple factors to obtain the image data initial model, and then the original image and the double-input CNN model are used for testing and optimizing the image data initial model to obtain the image data final model, so that multiple modeling is carried out based on the image based on multiple factors; according to the method, the image is accurately recognized by combining the double-input CNN model and fusing the multi-dimensional features to improve the robustness of judgment, and the final model of the image data can be continuously optimized along with the increase of recognized images, so that the image type can be accurately recognized under the condition that the moire is not obvious (or only depends on a single moire feature).

Description

Technical Field

[0001] The present invention belongs to the technical field of digital image processing, and in particular relates to a method for obtaining a moiré recognition image based on multiple modeling of multi-style images. Background Art

[0002] With the widespread use of digital images, there is an increasing need to distinguish between native images and screen-shot images. Native images are first-hand digital images, while screen-shot images are re-shot versions. There are essential differences between the two in generation paths and quality.

[0003] Currently, images captured by screens often contain moiré patterns, a visual artifact caused by a mismatch between the screen pixel grid and the camera sensor pixel grid. However, the presence of moiré patterns is not absolute, and their intensity and visibility depend on a variety of factors, such as shooting angle, distance, screen resolution, and camera settings. Therefore, directly using moiré patterns to distinguish between native images and screen-captured images will have the following technical drawbacks: 1. False negative / false positive risk False negatives (missed detections): When the shooting angle is vertical, the distance is far, and the screen resolution matches the camera sensor, the moiré pattern may disappear or be extremely subtle, causing the algorithm to be unable to recognize it.

[0004] False positives (false positives): Periodic textures in native images (such as stripes, grids, and regular patterns) can be easily mistaken for moiré patterns. 2. Image post-processing interference Compression damage characteristics: JPEG compression blurs the high-frequency details of moiré patterns; Geometric transformation interference: Rotation and scaling will change the direction and frequency of moiré patterns, making frequency domain analysis ineffective; Filter interference: Noise reduction and sharpening may eliminate or artificially add artifacts such as moiré; 3. Environmental dynamics Lighting and reflections: Strong reflections or low-light environments can mask moiré patterns. Motion blur: When shooting a moving screen (such as a video), motion blur destroys the periodic structure of the moiré pattern; 4. Data and model limitations Lack of generalization: The model needs to cover all screen types (such as Mini-LED and foldable screens) and camera device combinations, and the data collection cost is extremely high; Labeling dependency: A large number of screen images need to be manually labeled, and they need to be continuously updated as screen technology evolves. 5. Computational efficiency bottleneck Poor real-time performance: Frequency domain analysis (FFT) and deep learning models (such as CNN) experience significant latency on low-computing devices (such as mobile phones); 6. Insufficient algorithm robustness Local dependency flaw: only focusing on local textures (such as LBP) and ignoring global context such as screen borders and icons; Multi-scale problem: Moiré patterns appear very different at different scaling ratios, requiring multi-scale feature fusion.

[0005] Therefore, the reliability of relying on a single moiré feature detection is low. That is, the market urgently needs a method that can accurately identify screen-shot images and native photos even when the moiré patterns are not obvious, and this method integrates multi-dimensional features to improve the robustness of judgment. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a new method for obtaining moiré recognition images based on multiple modeling of multi-style images.

[0007] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows: A method for obtaining a moiré recognition image based on multiple modeling of multiple-style images comprises the following steps: S1. Image Multiple Modeling 1) M screen-shot images and N non-screen-shot images are collected and formed into data sets respectively, where each data set is divided based on factors such as screen type, resolution and refresh rate, angle and distance, lighting conditions, background pattern, dynamic and static, and shooting equipment, and is divided into training set, validation set, and test set; 2) LBP feature extraction is used to repeatedly train, validate, and test each data set, and the resulting moiré pattern is used as a benchmark to form an image data preliminary model; 3) The image data preliminary model is then tested and optimized using the native method. The LBP algorithm is first used for texture feature extraction, and then a dual-input CNN model is constructed. Feature fusion and SVM classification are then performed to optimize the image data preliminary model to complete the multi-modeling image data final model, where the dual-input CNN model includes a native image input stream and an LBP feature input stream. The native image is directly input into the native image input stream, and feature extraction is performed through convolution layers, pooling layers, and activation functions; a single-channel feature map generated by the LBP algorithm is input into the LBP feature input stream, and features are also extracted through a set of convolution layers, pooling layers, and activation functions; S2. Image Recognition Based on the above-mentioned multiple modeling, the image to be identified is input into the image data final model to obtain the moiré information of the image, and then the output value obtained based on the image data final model is compared with the decision threshold to identify the type of the input image.

[0008] Preferably, in step S1 1), the screen type includes LCD screen, OLED screen, LED screen; the resolution includes HD, FHD, 4K, 8K; the refresh rate includes 60Hz, 120Hz; the angle includes 0°, 30°, 45°, 60°, 90°; the distance includes close distance, medium distance, and long distance; the lighting conditions include natural light, indoor light, backlight, and direct sunlight; the background pattern is to display different patterns and colors; dynamic and static include dynamic video clips and static images; the shooting equipment includes optical imaging equipment and electronic imaging equipment.

[0009] In some specific embodiments, LCD screens include monitors of different brands and models, laptop screens, tablet screens, etc. OLED screens include smartphones, high-end TV screens, etc. LED screens include large outdoor advertising screens, electronic billboards, etc. Angles and distances are used to simulate various shooting conditions that may occur in the real world. Different lighting conditions refer to shooting under different lighting environments, including natural light, indoor lighting, backlight, direct sunlight, etc., to observe the performance of moiré patterns under various lighting conditions. Diverse background patterns include solid colors, stripes, grids, complex graphics, etc., to test the performance of the algorithm when facing different visual textures. Dynamic content and static content include dynamic video clips and static images to evaluate the stability of the algorithm when processing moving images. Diversity of shooting equipment, such as using cameras and mobile phones of different brands and models for shooting, to cover the impact of different sensor characteristics and lens distortion on the performance of moiré patterns.

[0010] According to a specific implementation and preferred aspect of the present invention, in step S1 1), the number of training set images in each dataset accounts for 75% to 85% of the total number of images in the corresponding dataset; the number of validation set images accounts for 5% to 10% of the total number of images in the corresponding dataset; and the remaining images are test set images. The accuracy of the constructed model is improved based on training, validation, and testing.

[0011] Ideally, the number of validation and test images should be equal, with the training set comprising at least 80% of the total. Typically, the dataset is 80% training, 10% validation, and 10% test. Using proportionally proportional training, validation, and testing, the accuracy of the image data model is continuously improved.

[0012] According to another specific implementation and preferred aspect of the present invention, the LBP feature extraction method used in step S1 is as follows:

[0013] Where: P is the number of sampling points; R is the radius of the sampling point from the center pixel; g c (x) is the gray value of the center pixel; g i(x) is the gray value of the ith neighborhood pixel; s(.) is the sign function, 2 i It is a weight factor that assigns different binary weights to different neighborhood positions. It returns 1 when the input is greater than or equal to 0, otherwise it returns 0.

[0014] Preferably, the LBP value can be calculated for each pixel, and the LBP feature map of the entire image can be obtained.

[0015] Furthermore, in step S1, the fusion mechanism fuses the dual-input CNN model features to distinguish weak moiré patterns and complex interference patterns in the screen capture image.

[0016] In some embodiments, the fusion mechanism includes splicing fusion and weighted fusion. The formula of splicing fusion is: f fusion=Wf×[ f fraw; f LBP]+bf in: f fraw is the feature vector of the original image stream; f LBP is the texture feature vector extracted based on the local binary pattern LBP; Wf is the weight matrix of the fusion layer; bf is the bias term of the fusion layer; [;] is the vector splicing operator; f fusion is the fused joint feature vector.

[0017] The formula for weighted fusion is: f fusion= α ⋅ f CNN+ β ⋅ f LBP in, f fusion is the joint feature vector after fusion; α and β is the weight parameter of CNN semantic features and LBP texture features; f CNN is a high-level semantic feature extracted by convolutional neural network; f LBP is a texture feature vector extracted based on the local binary pattern LBP.

[0018] Furthermore, in step S1, the feature vector extracted by CNN is used as input and the corresponding label is used as output to train the SVM classifier, and the kernel function is used for optimization.

[0019] In addition, in step S2, if the output value is less than the decision threshold, it is determined to be a native image, otherwise it is a screen-captured image.

[0020] Furthermore, manual recognition is performed based on the results obtained from the final model of the image data, and the decision threshold is corrected online during the manual recognition.

[0021] Due to the implementation of the above technical solution, the present invention has the following advantages compared with the prior art: The existing moiré pattern recognition algorithm for distinguishing native images from screen-captured images is subject to false negatives (missed detections): when the camera is shot at a vertical angle, at a long distance, or when the screen resolution matches the camera sensor, the moiré pattern may disappear or be extremely subtle, rendering the algorithm unable to recognize it. False positives (false positives): Periodic textures in native images (such as stripes, grids, and regular patterns) can be easily mistaken for moiré patterns. Compression damages features: JPEG compression blurs the high-frequency details of moiré patterns. Geometric transformation interference: Rotation and scaling change the direction and frequency of moiré patterns, rendering frequency domain analysis ineffective. Filter interference: Noise reduction and sharpening may eliminate or artificially add moiré-like artifacts. Lighting and reflections: Strong reflections or low-light environments can mask moiré patterns. Motion blur: When shooting a moving screen (such as a video), motion blur destroys the periodic structure of the moiré pattern; Lack of generalization: The model needs to cover all screen types (such as Mini-LED, folding screens) and camera equipment combinations, and the data collection cost is extremely high; Labeling dependency: A large number of screen shooting images need to be manually labeled, and they need to be continuously updated with the iteration of screen technology; Poor real-time performance: Frequency domain analysis (FFT) and deep learning models (such as CNN) have significant delays on low-computing power devices (mobile phones); Local dependency defects: Only focus on local textures (such as LBP), ignoring global context such as screen borders and icons; Multi-scale problems: Moiré patterns have large differences in performance at different scaling ratios, requiring multi-scale feature fusion and other deficiencies. The present invention comprehensively designs a method for obtaining moiré recognition images based on multiple modeling of multi-style images, and cleverly solves the various deficiencies of the existing structure. After adopting the method of obtaining moiré recognition images using multiple modeling, first, M screen-shot images and N non-screen-shot images are collected and formed into data sets respectively, where each data set is divided based on factors such as screen type, resolution and refresh rate, angle and distance, lighting conditions, background pattern, dynamic and static, and shooting equipment, and is divided into training set, validation set, and test set; secondly, LBP feature extraction is used to repeatedly train, validate and test each data set, and the obtained moiré pattern is used as a benchmark to form an image data preliminary model; again, the original image data preliminary model is tested and optimized respectively, first using the LBP algorithm for texture feature extraction, and then a dual-input CNN model is constructed, followed by feature fusion and SVM classification to optimize the image data preliminary model to complete the multi-modeling image data final model, and finally the image to be identified is input into the image data final model to obtain the image moiré information, and then the output value obtained based on the image data final model is compared with the decision threshold to identify the type of the input image.Therefore, the present invention performs multiple and repeated training, verification, and testing on images under multiple factors to obtain an initial model of image data, and then tests and optimizes the initial model of image data using the native image and the dual-input CNN model to obtain a final model of image data. Therefore, not only is multiple modeling performed on images based on multiple factors, but the dual-input CNN model and the fusion of multi-dimensional features are combined to improve the robustness of judgment to accurately identify images, and as the number of recognized pictures increases, the final model of image data can be continuously optimized, so that even when the moiré pattern is not obvious (or only relying on a single moiré pattern feature), the image type can be accurately identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flowchart of the multi-modeling moiré recognition method for multiple-style images of the present invention; Figure 2 for Figure 1 Test flow chart for multiple modeling in

[15] ; Figure 3 The following is a flow chart of image recognition for moiré pattern recognition of an image to be recognized. DETAILED DESCRIPTION

[0023] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The following description sets forth many specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar modifications without violating the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the present invention.

[0025] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, features specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0026] In the present invention, unless otherwise specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood in a broad sense. For example, they can refer to fixed connection, detachable connection, or integration; they can refer to direct connection or indirect connection through an intermediate medium; they can refer to internal communication between two components or interaction between two components, unless otherwise specified. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0027] In the present invention, unless otherwise expressly specified or limited, when a first feature is "above" or "below" a second feature, it may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediary. Furthermore, when a first feature is "above," "above," or "above" a second feature, it may mean that the first feature is directly above or diagonally above the second feature, or simply means that the first feature is at a higher level than the second feature. When a first feature is "below," "below," or "below" a second feature, it may mean that the first feature is directly below or diagonally below the second feature, or simply means that the first feature is at a lower level than the second feature.

[0028] It should be noted that when an element is referred to as being "fixed to" or "disposed on" another element, it may be directly on the other element or there may be an intermediate element. When an element is referred to as being "connected to" another element, it may be directly connected to the other element or there may be an intermediate element. The terms "vertical," "horizontal," "upper," "lower," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only implementation methods.

[0029] like Figure 1 and Figure 2 As shown, the method for obtaining a moiré recognition image based on multiple modeling of multi-style images involved in this embodiment includes the following steps: S1, image multiple modeling; S2, image recognition.

[0030] In some specific embodiments, image multiple modeling S1 includes the following steps: 1) collecting M screen-shot images and N non-screen-shot images, and forming data sets respectively, wherein each data set is divided based on factors such as screen type, resolution and refresh rate, angle and distance, lighting conditions, background pattern, dynamic and static, and shooting equipment, and is divided into training set, verification set, and test set; 2) using LBP feature extraction to repeatedly train, verify and test each data set, and form an image data initial model based on the obtained moiré pattern; 3) then using native to test and optimize the image data initial model, first using the LBP algorithm to extract texture features, and then constructing a dual-input CNN model, and then performing feature fusion and SVM classification to optimize the image data initial model to complete the image data final model of multiple modeling.

[0031] In this example, screen types include LCD screens, OLED screens, and LED screens; resolutions include HD, FHD, 4K, and 8K; refresh rates include 60Hz and 120Hz; angles include 0°, 30°, 45°, 60°, and 90°; distances include close, medium, and long distances; lighting conditions include natural light, indoor lighting, backlight, and direct sunlight; background patterns include displaying different patterns and colors; dynamic and static include dynamic video clips and static images; and shooting equipment includes optical imaging equipment and electronic imaging equipment.

[0032] In some specific embodiments, LCD screens include monitors of different brands and models, laptop screens, tablet screens, etc. OLED screens include smartphones, high-end TV screens, etc. LED screens include large outdoor advertising screens, electronic billboards, etc. Angles and distances are used to simulate various shooting conditions that may occur in the real world. Different lighting conditions refer to shooting under different lighting environments, including natural light, indoor lighting, backlight, direct sunlight, etc., to observe the performance of moiré patterns under various lighting conditions. Diverse background patterns include solid colors, stripes, grids, complex graphics, etc., to test the performance of the algorithm when facing different visual textures. Dynamic content and static content include dynamic video clips and static images to evaluate the stability of the algorithm when processing moving images. Diversity of shooting equipment, such as using cameras and mobile phones of different brands and models for shooting, to cover the impact of different sensor characteristics and lens distortion on the performance of moiré patterns.

[0033] The number of training images in each dataset accounts for 75% to 85% of the total number of images in the corresponding dataset; the number of validation images accounts for 5% to 10% of the total number of images in the corresponding dataset; and the remaining images are test images. The accuracy of the constructed model is improved through training, validation, and testing. In this example, the number of validation and test images is equal, and the number of training images accounts for at least 80%. Generally, 80% of the dataset is the training set, 10% is the validation set, and 10% is the test set. Using a proportional number of training, validation, and testing images, the accuracy of the constructed image data model is continuously improved.

[0034] In some specific embodiments, the dual-input CNN model includes a native image input stream and an LBP feature input stream. The native image is directly input into the native image input stream, and features are extracted through convolution layers, pooling layers, and activation functions; the single-channel feature map generated by the LBP algorithm is input into the LBP feature input stream, and features are also extracted through a set of convolution layers, pooling layers, and activation functions.

[0035] In this example, the LBP feature extraction method is used, and the formula is:

[0036] Where: P is the number of sampling points; R is the radius of the sampling point from the center pixel; g c (x) is the gray value of the center pixel; g i (x) is the gray value of the ith neighborhood pixel; s(.) is the sign function, 2 i is a weight factor that assigns different binary weights to different neighborhood positions. It returns 1 when the input is greater than or equal to 0, otherwise it returns 0. For each pixel, the LBP value can be calculated and the LBP feature map of the entire image can be obtained.

[0037] Furthermore, the fusion mechanism fuses the dual-input CNN model features to distinguish weak moiré patterns and complex interference patterns in the screen shot image. In some embodiments, the fusion mechanism includes splicing fusion and weighted fusion. The formula for splicing fusion is: f fusion=Wf×[ f fraw; f LBP]+bf in: f fraw is the feature vector of the original image stream; f LBP is the texture feature vector extracted based on the local binary pattern LBP; Wf is the weight matrix of the fusion layer; bf is the bias term of the fusion layer; [;] is the vector splicing operator; f fusion is the joint feature vector after fusion.

[0038] The formula for weighted fusion is: f fusion= α ⋅ f CNN+ β ⋅ f LBP in, f fusion is the joint feature vector after fusion; α and β is the weight parameter of CNN semantic features and LBP texture features; f CNN is a high-level semantic feature extracted by convolutional neural network; f LBP is a texture feature vector extracted based on the local binary pattern LBP.

[0039] The feature vectors extracted by the CNN are used as input, and the corresponding labels are used as output to train an SVM classifier, using a kernel function for optimization. Specifically, the appropriate kernel function, such as a linear kernel, a polynomial kernel, or a Gaussian kernel (RBF), is selected based on the performance of the validation set. SVM parameters, such as the penalty coefficient C and the kernel function parameter (e.g., γ), can also be adjusted to optimize classification performance.

[0040] In some specific embodiments, image recognition S2 involves inputting the image to be recognized into the final image data model based on the aforementioned multi-modeling process to obtain moiré information. This information is then compared against a decision threshold obtained from the final image data model to identify the input image type. In short, if the output value is less than the decision threshold, the image is considered native; otherwise, it is a screen capture. In this example, manual recognition is performed based on the results obtained from the final image data model, and the decision threshold is corrected online based on the manual recognition results. Therefore, based on manual re-inspection and online correction of the decision threshold based on the recognition results, the accuracy of the final image data model is further improved.

[0041] In summary, combined Figure 3 As shown, the implementation process of this embodiment is as follows: Image to be identified → LBP feature extraction → dual-input CNN model → splicing and fusion → SVM classification → decision threshold comparison → obtain image type.

[0042] At the same time, in order to further verify the accuracy of the image type, manual identification is performed based on the results obtained from the final model of the image data. In the manual identification, it is concluded that the image type result is correct and the identification is completed; otherwise, there is a deviation in the correct image type result, the decision threshold is corrected online, and then the decision threshold is corrected to continue the final model identification of the image data.

[0043] In summary, after adopting the multiple modeling to obtain the moiré recognition image method, first, M screen-shot images and N non-screen-shot images are collected, and data sets are formed respectively, where each data set is divided based on factors such as screen type, resolution and refresh rate, angle and distance, lighting conditions, background pattern, dynamic and static, and shooting equipment, and is divided into training set, verification set, and test set; secondly, LBP feature extraction is used to repeatedly train, verify and test each data set, and the obtained moiré is used as a benchmark to form an image data initial model; again, the original image data initial model is tested and optimized respectively, first using the LBP algorithm for texture feature extraction, and then constructing a dual-input CNN model, followed by feature fusion and SVM classification to optimize the image data initial model to complete the multiple modeling image data final model, and finally the image to be identified is input into the image data final model to obtain the moiré information of the image, and then the output value obtained based on the image data final model is compared with the decision threshold to identify the type of the input image, Therefore, on the one hand, the present invention performs multiple and repeated training, verification, and testing on images under multiple factors to obtain an initial model of image data, and then tests and optimizes the initial model of image data using the native image and the dual-input CNN model to obtain the final model of image data. Therefore, not only is multiple modeling performed on images based on multiple factors, but the dual-input CNN model and the fusion of multi-dimensional features are combined to improve the robustness of judgment to accurately identify images, but also, as the number of recognized pictures increases, the final model of image data can be continuously optimized, so that even when the moiré pattern is not obvious (or only relies on a single moiré pattern feature), the image type can be accurately identified, and the native image is determined based on the output value being less than the decision threshold, otherwise it is a screen-shot image; on the other hand, manual recognition is performed based on the results obtained from the final model of image data, and the decision threshold is corrected online in the results of manual recognition. Therefore, based on manual re-inspection, the decision threshold is corrected online according to the recognition result to further improve the accuracy of the final model of image data.

[0044] The above detailed description of the present invention is intended to enable persons familiar with the art to understand the contents of the present invention and implement them. It does not limit the scope of protection of the present invention. Any equivalent changes or modifications made based on the spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for obtaining a moiré recognition image based on multiple modeling of multiple-style images, characterized in that: It includes the following steps: S1. Image Multiple Modeling 1) M screen-shot images and N non-screen-shot images are collected and formed into data sets respectively, where each data set is divided based on factors such as screen type, resolution and refresh rate, angle and distance, lighting conditions, background pattern, dynamic and static, and shooting equipment, and is divided into training set, validation set, and test set; 2) LBP feature extraction is used to repeatedly train, validate, and test each data set, and the resulting moiré pattern is used as a benchmark to form an image data preliminary model; 3) The image data preliminary model is then tested and optimized using the native method. The LBP algorithm is first used for texture feature extraction, and then a dual-input CNN model is constructed. Feature fusion and SVM classification are then performed to optimize the image data preliminary model to complete the multi-modeled image data final model. The dual-input CNN model includes a native image input stream and an LBP feature input stream. The native image is directly input into the native image input stream, and feature extraction is performed through convolution layers, pooling layers, and activation functions; a single-channel feature map generated by the LBP algorithm is input into the LBP feature input stream, and features are also extracted through a set of convolution layers, pooling layers, and activation functions; S2. Image Recognition Based on the above-mentioned multiple modeling, the image to be identified is input into the image data final model to obtain the moiré information of the image, and then the output value obtained based on the image data final model is compared with the decision threshold to identify the type of the input image.

2. The method for obtaining a moiré recognition image based on multiple modeling of multiple-style images according to claim 1, characterized in that: In step S1 1), the screen type includes LCD screen, OLED screen, and LED screen; the resolution includes HD, FHD, 4K, and 8K; the refresh rate includes 60Hz and 120Hz; the angle includes 0°, 30°, 45°, 60°, and 90°; the distance includes close distance, medium distance, and long distance; the lighting conditions include natural light, indoor light, backlight, and direct sunlight; the background pattern is to display different patterns and colors; the dynamic and static include dynamic video clips and static images; and the shooting equipment includes optical imaging equipment and electronic imaging equipment.

3. The method for obtaining a moiré recognition image based on multiple modeling of multiple-style images according to claim 1 or 2, characterized in that: In step S1 1), the number of training set images in each data set accounts for 75% to 85% of the total number of images in the corresponding data set; the number of validation set images accounts for 5% to 10% of the total number of images in the corresponding data set; and the rest are test set images.

4. The method for obtaining a moiré recognition image based on multiple modeling of multiple-style images according to claim 3, characterized in that: The number of validation set images is equal to the number of test set images, and the number of training set images is at least 80%.

5. The method for obtaining a moiré recognition image based on multiple modeling of multiple-style images according to claim 1, characterized in that: The LBP feature extraction method used in step S1 is: Where: P is the number of sampling points; R is the radius of the sampling point from the center pixel; gc(x) is the gray value of the center pixel; gi(x) is the gray value of the i-th neighboring pixel; s(.) is the sign function, 2 i It is a weight factor that assigns different binary weights to different neighborhood positions. It returns 1 when the input is greater than or equal to 0, otherwise it returns 0.

6. The method for obtaining a moiré recognition image based on multiple modeling of multiple-style images according to claim 5, characterized in that: For each pixel, the LBP value can be calculated and the LBP feature map of the entire image can be obtained.

7. The method for obtaining a moiré recognition image based on multiple modeling of multiple-style images according to claim 1, characterized in that: In step S1, the fusion mechanism fuses the dual-input CNN model features to distinguish weak moiré patterns and complex interference patterns in the screen shot image.

8. The method for obtaining a moiré recognition image based on multiple modeling of multiple-style images according to claim 7, characterized in that: The fusion mechanism includes splicing fusion and weighted fusion. The formula of splicing fusion is: f fusion=Wf×[ f fear; f LBP]+bf in: f fraw is the feature vector of the original image stream; f LBP is the texture feature vector extracted based on the local binary pattern LBP; Wf is the weight matrix of the fusion layer; bf is the bias term of the fusion layer; [;] is the vector concatenation operator; f fusion is the joint feature vector after fusion; The formula for weighted fusion is: f fusion= α ⋅ f CNN+ β ⋅ f LBP;; in: f fusion is the joint feature vector after fusion; α and β is the weight parameter of CNN semantic features and LBP texture features; f CNN is a high-level semantic feature extracted by convolutional neural network; f LBP is a texture feature vector extracted based on the local binary pattern LBP.

9. The method for obtaining a moiré recognition image based on multiple modeling of multiple-style images according to claim 1, characterized in that: In step S1, the feature vector extracted by CNN is used as input and the corresponding label is used as output to train the SVM classifier, and the kernel function is used for optimization.

10. The method for obtaining a moiré recognition image based on multiple modeling of multiple-style images according to claim 1, characterized in that: In step S2, if the output value is less than the decision threshold, the image is determined to be a native image; otherwise, it is a screen capture image; and / or, manual recognition is performed based on the results obtained from the final model of the image data, and the decision threshold is corrected online during manual recognition.

Images