Recognition method and device of copied image, equipment and storage medium

Abstract

The invention provides a copied image recognition method, device and equipment and a storage medium, and relates to the technical field of computers, image processing, artificial intelligence, computer vision and the like. According to the specific implementation scheme, multiple types of image features are extracted from a target image; performing fusion operation on the multiple types of image features to obtain fusion features; and performing classification operation on the target image based on the fusion feature to obtain a copying recognition result of the target image.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to the fields of image processing, artificial intelligence, and computer vision. Background Technology

[0002] With the development of information technology, images, with their intuitiveness and ease of understanding, have become an important carrier in the information dissemination system, playing an irreplaceable role in many fields.

[0003] However, the widespread use of images has also given rise to a series of security risks. Among them, the misuse of images based on photocopying technology is becoming increasingly prominent, posing a serious threat to social order, information security, and individual rights. Therefore, the efficient detection of photocopied images has become an urgent need to safeguard information security and protect legitimate rights. Summary of the Invention

[0004] This disclosure provides a method, apparatus, device, and storage medium for recognizing reproduced images.

[0005] According to one aspect of this disclosure, a method for recognizing reproduced images is provided, comprising: Extracting multiple types of image features from the target image; Multiple types of image features are fused to obtain fused features; The target image is classified based on the fusion features to obtain the re-image recognition result of the target image.

[0006] According to another aspect of this disclosure, a device for recognizing reproduced images is provided, comprising: The extraction module is used to extract various types of image features from the target image; The fusion module is used to fuse multiple types of image features to obtain fused features; The recognition module is used to classify the target image based on fused features to obtain the re-image recognition result of the target image.

[0007] According to another aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and The memory is communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the methods described in the present disclosure.

[0008] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform any of the methods according to embodiments of this disclosure.

[0009] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements any of the methods according to embodiments of this disclosure.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a schematic flowchart of a method for recognizing reproduced images according to an embodiment of the present disclosure; Figure 2 This is a schematic diagram of a process for extracting local texture features according to an embodiment of the present disclosure; Figure 3 This is a flowchart illustrating the process of determining adaptive thresholds corresponding to each scale according to an embodiment of the present disclosure. Figure 4 This is a schematic flowchart illustrating the process of obtaining intermediate texture features at multiple scales according to an embodiment of the present disclosure. Figure 5 This is a flowchart illustrating the extraction of global semantic features according to an embodiment of the present disclosure; Figure 6 This is a schematic diagram of the process for extracting frequency domain features according to an embodiment of the present disclosure; Figure 7 This is a schematic diagram of the network structure of a method for recognizing reproduced images according to an embodiment of the present disclosure; Figure 8 This is a schematic diagram of the structure of a photographic image recognition device according to an embodiment of the present disclosure; Figure 9 This is a block diagram of an electronic device used to implement the image recognition method of the present disclosure embodiments. Detailed Implementation

[0012] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] The terms “first,” “second,” etc., used in this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion, such as including a series of steps or units. A method, system, product, or apparatus is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0014] It should be noted that, unless it is explicitly stated that there is a sequential order of execution between different operations, or that there is a sequential order of execution between different operations in terms of technical implementation, the execution order between multiple operations may not be significant, and multiple operations may be executed simultaneously.

[0015] In related technologies, the detection of reproduced images includes the following three categories: (1) Traditional image forensics methods based on handcrafted features; (2) End-to-end detection method based on deep learning; (3) Hybrid approach that combines the advantages of two types of technologies.

[0016] Among these methods, various approaches are used to extract essential features from reproduced images, and all demonstrate certain detection capabilities in specific scenarios. However, they also have significant limitations. For example, most methods suffer from single-modal limitations, and the detection performance of reproduced images needs improvement.

[0017] In view of this, this disclosure provides a method for recognizing reproduced images, which improves the detection efficiency of reproduced images by fusing multiple types of image features.

[0018] like Figure 1 The diagram shown is a flowchart illustrating the image recognition method provided in this disclosure, including the following: S101 extracts various types of image features from the target image.

[0019] The target image is any image that is to be re-photographed.

[0020] Extracting multiple types of image features from a target image refers to extracting information representations from the target image that can describe different dimensions of the image.

[0021] Various types of image features can include image features acquired in different ways, such as image features extracted using neural networks and image features extracted using traditional image feature extraction operators.

[0022] Multiple types of image features can also include the extraction of image features from different modalities. For example, spatial domain modal image features and frequency domain modal image features are different types of image features.

[0023] S102, perform a fusion operation on multiple types of image features to obtain fused features.

[0024] In practice, feature fusion can be achieved through methods such as splicing, summation, or by using a cross-attention mechanism to fuse multiple types of image features to obtain fused features.

[0025] S103, classify the target image based on the fusion features to obtain the re-photographing recognition result of the target image.

[0026] In practice, a classification head based on neural network layers can be used to classify the fused features and output the re-photographing recognition result of the target image, that is, to classify the target image into a re-photographed image and a non-re-photographed image.

[0027] In this embodiment, by extracting multiple types of image features from the target image, it is possible to capture information from different aspects of the image more comprehensively compared to related technologies. Furthermore, fusing these features integrates the advantages of each type, forming a more representative and discriminative fused feature. Finally, based on this fused feature, the target image is classified to obtain the re-photographing recognition result. This effectively improves the detection accuracy of re-photographed images in complex re-photographing scenarios (such as multi-light source shooting and high-resolution screen re-photographing), and minimizes the problem of related technologies failing to fully characterize the multi-dimensional features of re-photographed images.

[0028] In this embodiment of the disclosure, multiple types of image features can be extracted from the target image by performing at least two of the following operations: (1) Extract local texture features from the target image; Local texture features refer to the visual characteristics exhibited by the local arrangement patterns of pixels in an image. They reflect recurring, regular, or random patterns in the image. In this embodiment, local texture features mainly include screen moiré patterns and pixelation characteristics.

[0029] Pixelation is characterized by the appearance of obvious pixel blocks in an image. This is especially noticeable when reproducing images displayed on a screen. In the reproduced images, the content on the screen appears as obvious pixel blocks, which are usually larger than the actual image content.

[0030] Pixelation can also manifest as uneven color and jagged edges, especially in reproduced images.

[0031] (2) Extract global semantic features from the target image; Global semantic features refer to features in an image that reflect the overall content and semantic information. Unlike local features (such as edges, textures, corners, etc.), global semantic features focus on the overall meaning of the image, rather than just local visual patterns or structures.

[0032] Global semantic features can be used to capture global semantic inconsistencies in target images, such as inconsistencies in content caused by display device edge deformation or ambient light reflection.

[0033] (3) Extract frequency domain features from the target image.

[0034] Frequency domain features are the characteristics representing the distribution of frequency components of a target image after it has been transformed from the spatial domain to the frequency domain using a Fourier transform. For example, low-frequency components typically correspond to the overall outline and slowly changing parts of the target image, while high-frequency components correspond to the details and edge information of the target image.

[0035] In this embodiment, local texture features can capture fine-grained local features such as moiré patterns and pixel blocks appearing in the target image; global semantic features can capture global semantic inconsistencies in the target image; and frequency domain features can reveal the distribution characteristics of the target image at different frequencies. By comprehensively extracting these different types of image features, the target image can be described more comprehensively, the feature expression of the target image can be enriched, thereby improving the accuracy and reliability of identifying whether the target image is a reproduced image.

[0036] Because the color representation of color images is easily affected by factors such as light intensity, light color temperature, and shooting equipment parameters, the RGB values ​​of the same texture area can vary significantly under different color conditions. Therefore, when the target image is a color image, to avoid increasing computational complexity and introducing redundant interference due to color, the color image can be converted to a grayscale image first.

[0037] In practice, converting the target image to a grayscale image can be described by expression (1): (1) In the expression (), This represents the pixel value at coordinates (x, y) of the target image to be converted to grayscale; This represents the pixel value of the target image in the red channel; This represents the pixel value of the target image in the green channel; This represents the pixel value of the target image in the blue channel; 0.299, 0.587, and 0.114 are the weighting coefficients for the red, green, and blue channels, respectively.

[0038] After converting the target image to grayscale, local texture features are extracted from the target image, such as... Figure 2 As shown, it may include the following: S201, perform sampling operation on the target image to obtain intermediate images of the target image at multiple scales.

[0039] In practice, bilinear interpolation can be used to downsample the target image to obtain intermediate images at multiple scales. Different scales correspond to different resolutions.

[0040] S202, extract the initial texture features from the intermediate images at each scale.

[0041] S203, input the initial texture features corresponding to each scale into the corresponding convolutional neural network to obtain intermediate texture features at multiple scales.

[0042] S204 performs a fusion operation on intermediate texture features at multiple scales to obtain local texture features.

[0043] In this embodiment, multi-scale sampling is performed on the target image to obtain intermediate images at different scales. This covers the texture details of the target image at different resolutions, avoiding the omission of local texture features by a single scale. Extracting initial texture features from the intermediate images at each scale captures basic feature information such as edges, pixels, and texture patterns at the corresponding resolution. Inputting the initial texture features at each scale into the corresponding convolutional neural network yields intermediate texture features. The network's hierarchical feature extraction capability enhances the representational ability of the initial texture features. Fusion operations on the multi-scale intermediate texture features integrate texture information from different resolutions, making the obtained local texture features better represent the characteristics of the reproduced image, thereby improving the recognition accuracy of the reproduced image.

[0044] In this embodiment of the disclosure, intermediate images at multiple scales (i.e., greater than or equal to two scales) can be obtained.

[0045] In practice, in order to extract initial texture images with overall representation capabilities from details to local textures at different scales, step S201 involves sampling the target image to obtain intermediate images of the target image at multiple scales. This can be implemented as follows: The target image is downsampled using a first downsampling rate and a second downsampling rate to obtain intermediate images corresponding to the first downsampling rate and the second downsampling rate, respectively; the first downsampling rate is less than the second downsampling rate; the intermediate images at multiple scales also include the target image.

[0046] Downsampling is an operation in digital image processing that reduces the resolution of an image. Essentially, it involves selecting pixels from the original image at intervals (e.g., selecting pixels row-by-row / column-by-column) or locally pooling them (e.g., mean pooling, max pooling) to obtain a smaller image. The downsampling rate is the percentage reduction in the number of pixels in the horizontal and vertical directions of the image during downsampling.

[0047] For example, the first downsampling rate can be set to 2, which reduces the width and height of the image to half their original size. The second downsampling rate can be set to 4, which reduces the width and height of the image to one-quarter of their original size. This yields intermediate images corresponding to the first downsampling rate and the second downsampling rate.

[0048] In addition to intermediate images obtained through different downsampling rates, the target image itself can be used as one of the intermediate images at multiple scales because it has the original resolution and the most complete detail information.

[0049] In this embodiment, a first downsampling rate is used to perform downsampling operations to obtain corresponding intermediate images. This preserves most of the fine-grained texture features while avoiding excessive loss of high-frequency information. A second downsampling rate is used to perform downsampling operations to obtain corresponding intermediate images, further highlighting the global structure and overall contour features of the target image. Using the original-scale target image as an intermediate image also provides complete, unreduced detail information as a reference. By combining intermediate images at multiple scales, the system can more comprehensively capture the different levels of reproduction features of the target image at various scales, from the original high resolution to different low resolutions.

[0050] Regarding step S202 mentioned above, extracting initial texture features from the intermediate images at each scale can be implemented as follows: An adaptive threshold is used to extract local binary pattern features from the intermediate images at each scale, which are then used as initial texture features.

[0051] Local Binary Pattern (LBP) features are a type of feature in computer vision used to describe the local texture of an image. Essentially, it compares the grayscale values ​​of a central pixel with those of its predefined neighborhood sampling points, representing the comparison result in binary form to form a local binary pattern. This binary pattern can be used to characterize the texture information of local regions of a target image, such as the smoothness, roughness, and regularity of the image.

[0052] In implementation, for each scale of the intermediate image, each pixel can be used as a center pixel. For each center pixel, its grayscale value is first compared with the grayscale values ​​of other sampling points within a preset neighborhood. If the grayscale value of a sampling point within the preset neighborhood is greater than or equal to the grayscale value of the center pixel, it is recorded as 1; otherwise, it is recorded as 0. These comparison results are then arranged into a binary number according to the sampling order, and this binary number is converted into a decimal number to obtain the local binary pattern features at each scale. The preset neighborhood is a fixed-size range defined centered on the center pixel, containing multiple pixels adjacent to the center pixel.

[0053] During implementation, the local binary pattern features at each scale are obtained, which can be described by expressions (2) and (3): (2) In expression (2), This represents the number of pixels in the preset neighborhood of the center pixel at scale s. The binary bits corresponding to each sampling point; This indicates that at scale s, the first in the preset neighborhood is... The grayscale value of each sampling point; Indicates the coordinates of the sampling point; This represents the adaptive threshold for the center pixel at scale s; s represents the coordinates of the center pixel; s represents the resolution scale of the target image.

[0054] (3) In expression (3), This represents the local binary pattern feature at the center pixel at scale s; Indicates to Sum of the binary bits of each sampling point; Indicates the first The weight of each binary bit is used to convert a binary string into an integer; Indicates the number of sampling points; Indicates the sampling point number within the preset neighborhood. =0,1,…, -1.

[0055] In this embodiment, an adaptive threshold is used to extract local binary pattern features from intermediate images at each scale as initial texture features. The adaptive threshold can dynamically adjust the threshold parameter according to the gray-level distribution characteristics of different regions of the image, effectively avoiding feature extraction deviations under complex lighting and contrast variation scenarios with a fixed threshold, and improving the accuracy of obtaining initial texture features.

[0056] For each scale of the intermediate image, an adaptive threshold corresponding to that scale is determined, such as... Figure 3 As shown, it includes the following: S301, for each center pixel in the intermediate image at this scale, determine the standard deviation and mean of the pixel values ​​in the preset neighborhood of the center pixel.

[0057] During implementation, the scale of the preset neighborhood can be determined according to the requirements.

[0058] The standard deviation of pixel values ​​within a preset neighborhood of the center pixel reflects the dispersion of pixel values ​​within that neighborhood. A larger standard deviation indicates more drastic changes in pixel values ​​within the preset neighborhood, potentially indicating the presence of obvious texture or edge information; a smaller standard deviation indicates more uniform pixel values ​​within the neighborhood.

[0059] The mean value of the pixel values ​​in the preset neighborhood of the center pixel represents the average level of the pixel values ​​in the preset neighborhood.

[0060] S302, based on the weight coefficients corresponding to this scale, the standard deviation, the mean, and the pixel value of the center pixel are weighted and summed to obtain the adaptive threshold corresponding to this scale.

[0061] In implementation, at the original scale, subtle texture changes can be captured by strengthening the standard deviation weight, while also considering the center pixel value and local mean to achieve high-frequency detail enhancement. In terms of weight allocation, the standard deviation weight can be set to the highest, while the weights of the center pixel value and the mean can be set to almost equal (e.g., the difference between them is less than a preset difference), and both weights can be slightly lower than the annotation difference weight. For example, the weight corresponding to the center pixel value... =0.3; the weight corresponding to the mean. =0.3; the weight corresponding to the standard deviation =0.4.

[0062] At low to mid-frequency scales, such as the 1 / 2 scale, noise can be suppressed by reducing the standard deviation weight, while the weight of the center pixel can be increased to adapt to image smoothing effects, thus balancing the extraction of low to mid-frequency features. In terms of weight allocation, the pixel value weight of the center pixel can be set to the highest, followed by the mean weight, and the standard deviation weight to the lowest. For example, the weight corresponding to the pixel value of the center pixel... =0.5; the weight corresponding to the mean. =0.4; the weight corresponding to the standard deviation =0.1.

[0063] At low to mid-frequency scales, such as the 1 / 4 scale, texture interference can be minimized by removing the standard deviation weight, while maximizing the pixel value weight of the center pixel to enhance subject features and adapt to the smoothed image after significant downsampling. In terms of weight allocation, the pixel value weight of the center pixel can be set to the highest, followed by the mean weight, and the standard deviation weight can be zero. For example, the weight corresponding to the pixel value of the center pixel... =0.7; the weight corresponding to the mean. =0.3; the weight corresponding to the standard deviation =0.

[0064] In practice, based on the weight coefficients corresponding to the scale, the standard deviation, mean, and pixel value of the center pixel are weighted and summed to obtain the adaptive threshold corresponding to the scale, which can be described by expression (4): (4) In expression (4), This represents the adaptive threshold of the center pixel at scale s; This represents the grayscale value of the center pixel. The average gray value of pixels within a preset neighborhood of the center pixel; The grayscale standard deviation of pixel values ​​within a preset neighborhood of the center pixel; , , These are the weighting coefficients for each scale dependency.

[0065] In this embodiment, for each scale of the intermediate image, the standard deviation and mean of the pixel values ​​within a preset neighborhood of each center pixel are calculated to quantify the grayscale fluctuation of the pixels in that neighborhood and the overall grayscale level. Then, by combining the weighting coefficients corresponding to that scale, the standard deviation, mean, and center pixel values ​​are weighted and summed to determine an adaptive threshold. This effectively overcomes the oversegmentation or undersegmentation problems that easily occur in multi-scale image feature extraction using fixed thresholds, thereby improving the accuracy of subsequent local texture feature extraction.

[0066] The preset neighborhood of the center pixel can be set as a circular neighborhood. For the intermediate image at each scale, the neighborhood of the center pixel is... At this point, the sampling point within the circular neighborhood can be defined by expression (5): (5) In expression (5), Represents the first element in the preset neighborhood (circular neighborhood). The coordinates of each sampling point; R represents the coordinates of the center pixel; R represents the sampling radius. Indicates the first in the preset neighborhood The angle of each sampling point; , The trigonometric function representing the conversion from polar coordinates to rectangular coordinates is used to calculate the horizontal and vertical offsets of sampling points within a preset neighborhood relative to the center pixel.

[0067] Specifically, the sampling radius of the circular neighborhood used to extract local binary pattern features increases as the scale decreases. At low scales, image resolution decreases and detail information is compressed; increasing the sampling radius covers a larger pixel area, thus compensating for the loss of image detail at low scales. At high scales, image resolution is high and fine-grained textures are rich; using a smaller sampling radius can accurately capture subtle local texture structures and edge features. Therefore, by setting the sampling radius of the circular neighborhood for extracting local binary pattern features to increase as the image scale decreases, discriminative initial texture features can be extracted from intermediate images at different scales.

[0068] In step S203, the initial texture features corresponding to each scale are input into the corresponding convolutional neural network to obtain intermediate texture features at multiple scales. For example... Figure 4 As shown, the following operations are performed for the initial texture features at each scale: S401, an encoder based on a scale-corresponding convolutional neural network encodes the initial texture features to obtain encoded features; wherein, the encoder uses multiple types of convolutional kernels to process the initial texture features; different types of convolutional kernels are used to extract re-image artifact features in different directions.

[0069] S402, based on the scale-corresponding convolutional neural network decoder, processes the encoded features to obtain the intermediate texture features of the scale.

[0070] In this context, the decoder of the convolutional neural network can serve as a purifier. It is used to filter, enhance, and reconstruct the encoded features output by the encoder, retaining the re-photographing features in the target image and removing redundant noise and irrelevant information to reduce the complexity of subsequent processing of intermediate texture features.

[0071] In this embodiment, an encoder of a corresponding convolutional neural network performs an encoding operation to obtain encoded features. The encoder uses multiple types of convolutional kernels to process the initial texture features, enabling the capture of periodic artifacts in different directions in the image. The encoded features are then processed by a corresponding decoder to obtain intermediate texture features at that scale. This enhances the ability of the encoded features to represent artifacts at different scales, resulting in multi-scale intermediate texture features with stronger adaptability for re-image detection.

[0072] The encoder employs multiple types of convolutional kernels, including at least two of the following: (1) Horizontal stripe detection kernel, used to detect periodic stripes in the horizontal direction; In practice, the horizontal stripe detection kernel can be described by expression (6): (6) In expression (6), This represents the kernel element value at the i-th row and j-th column position of the horizontal stripe detection kernel; H represents the horizontal direction. The horizontal sinusoidal component is used to match periodic stripes in the horizontal direction; i represents the row index of the horizontal stripe detection kernel, corresponding to the vertical direction of the target image. This represents the estimated wavelength of the horizontal stripes, which is also the period of the horizontal stripes. The larger the size, the sparser the horizontal stripes; The phase term represents the sinusoidal component in the horizontal fringe detection kernel, used to control the periodic variation of the sine wave to match the horizontal stripes. represents the Gaussian window function, used to weight kernel elements, giving higher weights to the center of the kernel and lower weights to the edges, thereby reducing edge artifacts; j represents the column index of the horizontal stripe detection kernel, corresponding to the horizontal direction of the image; c represents the position of the center pixel. This represents the standard deviation of the Gaussian function and is used to control the width of the Gaussian window. The smaller the value, the more concentrated the effective range of the horizontal stripe detection kernel is at the center pixel.

[0073] (2) Vertical stripe detection kernel, used to detect periodic stripes in the vertical direction; In practice, the vertical stripe detection kernel can be described by expression (7): (7) In expression (7), V represents the kernel element value of the vertical stripe detection kernel at the position of the i-th row and j-th column; V represents the vertical direction. Represents the sinusoidal component in the vertical direction; This indicates the estimation of the vertical fringe wavelength; This represents the phase term of the sinusoidal component in the vertical fringe detection kernel; This represents the Gaussian window function; the other parameters are explained above and will not be repeated here.

[0074] (3) Grid detection kernel, used to detect grid features; In practice, the mesh detection kernel can be described by expression (8): (8) In expression (8), G represents the kernel element value at the i-th row and j-th column of the mesh detection kernel; G represents the mesh. Represents the sinusoidal component in the vertical direction; This represents the wavelength estimate of the vertical stripes. The larger the size, the sparser the vertical stripes. Represents the sinusoidal component in the horizontal direction; This represents the wavelength estimate of the horizontal stripes. The larger the size, the sparser the horizontal stripes.

[0075] (4) Diagonal cross detection kernel, used to detect periodic textures in the diagonal direction.

[0076] In practice, the diagonal crossover detection kernel can be described by expression (9): (9) In expression (9), This represents the kernel element value at the i-th row and j-th column of the diagonal cross-detection kernel; D represents the diagonal. Indicates diagonal stripes running from the top left to the bottom right; This represents the wavelength estimate of the diagonal fringes. The larger the size, the sparser the diagonal stripes; Indicates the periodic change in the direction of the backslash "\"; This indicates the diagonal stripes running from the upper right to the lower left.

[0077] In this embodiment, the horizontal stripe detection kernel can specifically capture periodic stripe features in the horizontal direction of the image. The vertical stripe detection kernel can accurately identify periodic stripe information in the vertical direction. The grid detection kernel can effectively extract grid-like texture features formed by the interweaving of horizontal and vertical stripes. The diagonal cross detection kernel can efficiently detect diagonally distributed periodic textures. By integrating the detection capabilities of at least two of the above-mentioned convolutional kernels, periodic textures in different directions in the image can be covered, effectively compensating for the problem of missed detection of texture features by single-direction convolutional kernels, thereby enhancing the ability to identify periodic artifacts such as moiré patterns and scan lines that are prone to appear in reproduced images, and providing more discriminative texture features for the subsequent recognition of reproduced images.

[0078] In step S401, the encoder of the scale-corresponding convolutional neural network encodes the initial texture features to obtain encoded features, which can be achieved based on the following steps: Step A1: The initial texture features are processed by multiple first convolutional blocks in the encoder to obtain the first intermediate features; wherein, multiple types of convolutional kernels are located in the first convolutional layer, and each type of convolutional kernel performs convolution operation on the initial texture features to construct channel features. During implementation, the horizontal stripe detection kernel, vertical stripe detection kernel, grid detection kernel, and diagonal cross detection kernel perform convolution operations on the initial texture features respectively, and the independent processing result of each type of convolution kernel corresponds to a channel feature.

[0079] Each first convolutional block comprises a concatenated convolutional layer, a normalization layer, and an activation layer. The output features of the previous first convolutional block are pooled and then input into the next first convolutional block. The convolutional layers extract spatial features and transform the initial texture features, uncovering local texture correlations. The first normalization layer calibrates the data distribution of the convolutional output features, eliminating internal covariate shifts and providing a stable feature distribution foundation for subsequent activation operations. The activation layer introduces nonlinear representation capabilities through nonlinear transformations, overcoming the limitations of linear models and enhancing the discriminative power of texture features. The second normalization layer further regularizes the feature distribution after activation, avoiding feature distribution shifts caused by the activation function, ensuring that features maintain a stable and discriminative distribution before entering the next first convolutional block. This extracts typical features that can effectively distinguish whether an image is a copy, improving the accuracy of copy image recognition.

[0080] In implementation, the initial texture features are processed by multiple concatenated first convolutional blocks in the encoder to obtain the first intermediate features, which can be described by expressions (10) and (11): (10) In expression (10), This represents the output features of the first convolutional block; Represents various types of convolution kernels for the first convolutional block; Indicates the convolution operation; Indicates the initial texture features; This represents the activation function, used to perform nonlinear mapping on the results of linear transformations to enhance feature representation capabilities; This indicates batch normalization.

[0081] (11) In expression (11), This represents the output features of the second first convolutional block; Represents various types of convolution kernels for the second first convolution block; This represents max pooling, used to downsample the intermediate features of the first convolutional block output while preserving key features and reducing computational cost.

[0082] Step A2: The first intermediate feature is processed using the second convolutional block in the encoder to obtain the encoded feature.

[0083] In practice, the first intermediate feature is processed by the second convolutional block in the encoder to obtain the encoded feature, which can be described by expression (12): (12) In expression (12), This represents the obtained encoded features; Represents various types of convolution kernels for the second convolutional block; This represents global average pooling, used to compress the normalized first intermediate feature from a two-dimensional feature map into a one-dimensional vector.

[0084] In this embodiment, multiple first convolutional blocks cascaded in the encoder process the initial texture features to obtain first intermediate features. Various types of convolutional kernels are used to perform convolution operations on the initial texture features to construct channel features, enabling the extraction of discriminative information from different dimensions and enriching the channel dimension representation capability of the features. The first intermediate features are further processed by second convolutional blocks in the encoder to obtain encoded features. The processing of the second convolutional blocks (such as hierarchical convolution, activation, and normalization operations) allows for feature fusion, dimensionality compression, and nonlinear mapping of the multi-channel first intermediate features, enhancing the feature representation capability. This makes the obtained encoded features more discriminative.

[0085] In step S402, the decoder of the scale-corresponding convolutional neural network processes the encoded features to obtain the intermediate texture features of the scale, which can be achieved based on the following steps: Step B1: For each scale, the encoded features are processed by a series of multi-layer fully connected networks in the decoder corresponding to the scale to obtain the second intermediate features; Taking the processing of encoded features by a two-layer fully connected network to obtain the second intermediate feature as an example: First, the encoded features are processed using a first-layer fully connected network, a process that can be described by expression (13): (13) In expression (13), This represents the output characteristics of the first fully connected layer network; This represents the weight matrix of the first fully connected layer of the network; The bias vector of the first fully connected layer; Represents encoded features; This indicates a linear transformation of the encoded features; This represents the activation function.

[0086] Next, a second fully connected network is used to process the output features of the first fully connected network to obtain the second intermediate features. This process can be described by expression (14): (14) In expression (14), This represents the output feature of the second fully connected layer, also known as the second intermediate feature; This represents the weight matrix of the second fully connected layer; The bias vector of the second fully connected network; This represents the output characteristics of the first fully connected layer network; This represents a linear transformation of the output features of the first fully connected layer network; This represents the activation function.

[0087] Step B2 involves processing the second intermediate feature using a linear transform layer in the encoder to obtain the intermediate texture feature at the scale.

[0088] In practice, the second intermediate feature is processed by the linear transformation layer in the encoder to obtain the intermediate texture features at the scale, which can be described by expression (15): (15) In expression (15), This represents the intermediate texture features obtained at each scale; This represents the weight matrix of the linear transform layer in the encoder. This represents the bias vector of the linear transform layer in the encoder. This indicates the second intermediate feature.

[0089] In this embodiment, for each scale, a multi-layer fully connected network cascaded in the corresponding decoder is used to process the encoded features to obtain the second intermediate features. The multi-layer fully connected network can restore the texture information in the encoded features to a feature expression that is more in line with the essence of the image texture. At the same time, the global feature integration capability of the fully connected layer can make up for the shortcomings of convolution operation in capturing global correlation information. The second intermediate features are processed by the linear transformation layer in the encoder to obtain the intermediate texture features of that scale. The linear transformation layer can perform dimensional adaptation and numerical calibration on the second intermediate features. The combination of these two steps achieves both deep decoupling and global integration of encoded features through the multi-layer fully connected network and accurate adaptation of the feature space through the linear transformation layer. The resulting intermediate texture features at each scale have both abstract representation capability and texture detail restoration, laying a high-quality foundation for subsequent multi-scale texture feature fusion.

[0090] In step S204, intermediate texture features from multiple scales are fused to obtain local texture features. This can be achieved through the following steps: Step C1: Based on the channel attention network, the splicing features of local texture features at various scales are processed to obtain attention weights; In practice, the local texture features at various scales are spliced ​​together to obtain spliced ​​features, which can be described by expression (16): (16) In expression (16), This represents the obtained splicing features; This represents the local texture features at the original scale (i.e., the texture features of the target image). Represents local texture features at a 1 / 2 scale; This represents a local texture feature at a 1 / 4 scale.

[0091] In practice, the attention weights are obtained by processing the spliced ​​features of local texture features at various scales based on the channel attention network, which can be described by expressions (17) and (18): (17) In expression (17), z represents the output feature of the first layer of the channel attention network; This represents the weight matrix of the first linear layer; This represents the bias matrix of the first linear layer; Indicates splicing features Perform a linear transformation; This represents the activation function.

[0092] (18) In expression (18), This indicates that the attention weights have been obtained; This represents the Sigmoid function, which compresses the output to the [0,1] interval and is used to represent the strength of attention weights; This represents the weight matrix of the second linear layer; This represents the bias matrix of the second linear layer; This indicates that a quadratic linear transformation is performed on the intermediate feature z.

[0093] Step C2 involves splitting the attention weights to obtain sub-weights; Step C3: Perform feature reshaping operation on each sub-weight to obtain the scale weights for each scale. In practice, the attention weights are split to obtain sub-weights, and the sub-weights are reshaped to obtain the scale weights for each scale, which can be described by expression (19): (19) In expression (19), represents the scale weights obtained at each scale; s represents the index corresponding to each sub-weight; a represents the attention weight to be split; d represents the target dimension of each sub-weight; This indicates a splitting operation on the attention weights: starting from the s·d-th element of attention weight a, take the (s+1)·d-1-th element to obtain a sub-weight of length d; This indicates a reshaping operation, used to reshape the individual sub-weights.

[0094] Step C4: Based on the weights of each scale, the intermediate texture features of each scale are weighted and fused to obtain weighted features; During implementation, the intermediate texture features at each scale are weighted and fused to obtain weighted features, which can be described by expression (20): (20) In expression (20), This represents the obtained weighted features; Indicates the original scale; Represents local texture features at the original scale; Indicates a 1 / 2 scale; Represents local texture features at a 1 / 2 scale; Indicates a 1 / 4 scale; Represents local texture features at a 1 / 4 scale; This represents element-wise multiplication; This indicates feature splicing.

[0095] Step C5 involves performing a feature transformation operation on the weighted features to obtain local texture features.

[0096] In practice, a feature transformation operation is performed on the weighted features to obtain local texture features, which can be described by expression (21): (twenty one) In expression (21), This represents the obtained local texture features; This represents the weight matrix of the first linear layer (coding layer); This represents the bias vector of the first linear layer; Indicates weighted features; This indicates a linear transformation applied to the weighted features; This represents an activation function that performs a nonlinear mapping on the linearly transformed features, thereby enhancing the expressive power of the features. This represents the weight matrix of the second linear layer (feature mapping layer); This represents the bias vector of the second linear layer; Indicates to The features output by the function are subjected to a quadratic linear transformation. The representation layer normalization operation is used to normalize the features after a quadratic linear transformation.

[0097] In this embodiment, attention weights are decomposed into sub-weights that match features at each scale, thus adapting the sub-weights to their corresponding scales. Feature reshaping of each sub-weight yields scale weights for each scale, adjusting their dimensions and form to match the feature dimensions of intermediate texture features at the corresponding scale. Weighted fusion of intermediate texture features at each scale weight yields weighted features, and feature transformation is performed on these weighted features to obtain local texture features. This enables adaptive and differentiated fusion of multi-scale texture features, providing highly discriminative local texture feature dimensions for the recognition of reproduced images, thereby improving the recognition accuracy of reproduced images.

[0098] In this embodiment of the disclosure, global semantic features are extracted from the target image, such as... Figure 5 As shown, it includes the following: S501, the target image is divided into multiple image blocks.

[0099] In implementation, a lightweight Vision Transformer (ViT) architecture can be used to segment the input image into multiple fixed-size image blocks.

[0100] S502, perform edge-preserving filtering on each image block to obtain the filtered features.

[0101] Edge-preserving filtering is a type of filtering algorithm whose main goal is to preserve edge information in an image while smoothing it.

[0102] In practice, edge-preserving filtering algorithms such as bilateral filtering and guided filtering can be used to perform edge-preserving filtering on each image block to obtain the filtered features.

[0103] Taking bilateral filtering as an example, edge-preserving filtering is performed on each image block to obtain the filtering features, which can be described by expression (22): (twenty two) In expression (22), The filtered features of the obtained image patch are represented; This represents the bilateral filtering algorithm; This represents the m-th image block obtained after segmenting the target image; This represents the standard deviation in the spatial domain, used to control the filtering range. This represents the standard deviation of the grayscale range, used to control the grayscale similarity weight and enhance the difference between abnormal edges and the background in re-enacted images.

[0104] S503 inputs the filtered features of each image block into the convolution processing module to obtain local features.

[0105] That is, after the filtered features of each image block are input into the convolution processing module, the convolution kernel in the convolution module performs a sliding convolution operation on the filtered features of each image block to obtain the local features of each image block.

[0106] During implementation, the filtered features of each image block are input into the convolution processing module to obtain local features, which can be described by expression (23): (twenty three) In expression (23), Represents the local features of the obtained image patch; This represents the convolution kernel in the convolution module; This is a convolution operation; Represents the filtering features of an image patch; Represented as convolution bias term; This represents the activation function, which restricts the output range to the interval [0,6] to avoid gradient explosion. This indicates batch normalization, used to stabilize feature distribution and improve model generalization ability.

[0107] S504 inputs the local features of each image patch into the encoder to obtain global semantic features.

[0108] In implementation, the local features of each image patch are first flattened into one-dimensional vectors and mapped through a linear projection layer to embedding vectors with uniform dimensions that conform to the input format of the ViT encoder. Then, a learnable CLS (Classification Token) is concatenated to the front of this embedding vector sequence. The concatenated sequence is then input into the ViT transformer encoder. The encoder models the long-distance dependencies between the embedding vectors of all image patches through a multi-layer self-attention mechanism, while the CLS token continuously aggregates global feature information, ultimately extracting the global semantic features of each image patch.

[0109] In this embodiment, by segmenting the target image into multiple image blocks, the overall semantics of the target image can be decomposed into semantic units of local regions, reducing the complexity of global semantic extraction while focusing on fine-grained semantic features of different regions. Edge-preserving filtering is performed on each image block to obtain filtered features, which can smooth redundant noise within the image block while fully preserving key semantically related structural information such as edges and contours. The filtered features of each image block are input into a convolution processing module to obtain local features. Convolution operations can be used to mine local spatial relationships and texture semantic information within the image block, enhancing the semantic representation capability of local regions. The local features of each image block are then input into an encoder to obtain global semantic features. The encoder can aggregate the semantic information of scattered local regions into a high-level semantic representation of the target image as a whole, effectively improving the completeness of global semantic features and providing a highly discriminative global semantic dimension for the recognition of reproduced images.

[0110] In this embodiment of the disclosure, frequency domain features are extracted from the target image, such as... Figure 6 As shown, it includes the following: S601, perform moiré pattern removal processing on the target image to obtain an optimized image.

[0111] Moiré patterns are a common type of interference pattern in images, caused by a mismatch between the sampling frequency of the imaging device and the texture frequency of the subject. In images, they typically appear as irregular, periodic stripes or patterns.

[0112] In practice, a low-pass filter can be used to remove moiré patterns from the target image to obtain an optimized image; alternatively, image analysis and processing algorithms, such as adaptive filtering and wavelet transform, can be used to selectively remove moiré patterns based on the characteristics of the target image to obtain an optimized image.

[0113] S602, transform the target image and the optimized image to the frequency domain respectively to obtain the first spectral image and the second spectral image.

[0114] In practice, the Fast Fourier Transform method can be used to map the target image and the optimized image from the spatial domain to the frequency domain, ultimately obtaining the first spectral image corresponding to the target image and the second spectral image corresponding to the optimized image.

[0115] S603, determine the difference between the first and second spectral images to obtain frequency domain features.

[0116] During implementation, the differences between the first and second spectral images are determined, and quantitative indicators such as corresponding pixel differences, mean square errors, and structural similarity indices are used to calculate the frequency domain features.

[0117] In this embodiment, an optimized image is obtained by performing moiré pattern removal processing on the target image, which effectively removes periodic artifact noise such as moiré patterns that are prone to appear in re-photographed images, avoiding interference with frequency domain feature extraction. The target image and the optimized image are transformed to the frequency domain to obtain a first spectral image and a second spectral image, respectively. The difference between the first spectral image and the second spectral image is calculated to obtain frequency domain features. By observing changes in these frequency domain features, frequency domain anomalies related to re-photographing in the target image can be captured, providing a highly discriminative frequency dimension basis for the recognition of re-photographed images.

[0118] In this embodiment of the disclosure, after calculating various types of image features such as local texture features, global semantic features, and frequency domain features, the various types of image features can be processed based on a multi-scale cross-attention fusion mechanism to obtain fused features; wherein, local texture features are used as query features, global semantic features are used as key features, and frequency domain features are used as value features.

[0119] In implementation, local texture features, global semantic features, and frequency domain features can be adjusted to the same channel scale first. Then, using local texture features as the query (Q), global semantic features as the key (K), and frequency domain features as the value (V), a multimodal cross-attention calculation module is constructed to quantify the semantic correlation between different feature modalities. Finally, based on the calculated correlation weights, the three types of features are weighted and fused to obtain the fused feature.

[0120] In this embodiment of the disclosure, by setting local texture features as query features, global semantic features as key features, and frequency domain features as value features, and processing local texture features, global semantic features, and frequency domain features based on a multi-scale cross-attention fusion mechanism to obtain fused features, the complementarity of the three types of features in the representation dimension can be achieved, effectively improving the discriminative power of the fused features for whether the target image is a reproduced image, and providing multi-dimensional feature support for reproduced images that includes details, semantics, and frequency information.

[0121] In summary, the overall structure of the image recognition network provided in this embodiment is as follows: Figure 7As shown, this re-photographed image recognition network can be obtained through supervised training using a high-value training sample set. Figure 7 As shown, the processing flow of the image recognition network for reproducing images includes the following: (1) Extracting local texture features: The target image is sampled to obtain intermediate images of the target image at multiple scales. Figure 7 Taking three scales as an example, each scale uses its own adaptive threshold to extract local binary pattern features from the intermediate images of that scale, serving as initial texture features. The convolutional neural network structure corresponding to each scale is identical (though the parameters may differ), extracting intermediate texture features from the initial texture features of each scale. Then, the intermediate texture features from each scale are fused together to obtain local texture features. Taking the processing of the initial texture features at scale 3 as an example, as follows... Figure 7 As shown, in the encoder 701 of the convolutional neural network, multiple cascaded first convolutional blocks 7011 process the initial texture features to obtain first intermediate features; the second convolutional block 7012 in the encoder processes the first intermediate features to obtain encoded features. For each scale, the encoded features are processed by a cascaded multi-layer fully connected network 7021 in the corresponding decoder 702 to obtain second intermediate features; the second intermediate features are then processed by a linear transform layer 7013 in the encoder 701 to obtain intermediate texture features for that scale. The intermediate texture features from multiple scales are then fused to obtain local texture features.

[0122] (2) Extract global semantic features: like Figure 7 As shown, a lightweight Vision Transformer architecture is employed to segment the input target image into multiple fixed-size image patches. Edge-preserving filtering is performed on each segmented image patch to obtain filtered features. Subsequently, the filtered features of each image patch are input into the convolution processing module 703 to obtain local features. Finally, the local features of each image patch are sequentially input into the transformer encoder 704 to obtain global semantic features.

[0123] (3) Extracting frequency domain features: like Figure 7 As shown, a low-pass filter 705 is first used to perform moiré pattern removal on the target image to obtain an optimized image. The target image and the optimized image are then mapped from the spatial domain to the frequency domain through frequency domain transformation, resulting in a first spectral image and a second spectral image. The differences between the first spectral image and the second spectral image are determined to obtain the frequency domain features.

[0124] (4) Integrating local texture features, global semantic features, and frequency domain features: like Figure 7As shown, a cross-attention mechanism 706 is used to fuse multiple types of image features to obtain fused features.

[0125] (5) Image recognition of reproduced images: like Figure 7 As shown, the fused features are sequentially processed through global average pooling, linear transformation and activation calculation of fully connected layer 707 and fully connected layer 708, and then probability normalization is completed through the softmax classification function to obtain the re-photographing recognition result of the target image.

[0126] In summary, the image recognition method for reproduced images provided by this disclosure not only improves the accuracy and environmental adaptability of reproduced image detection, but also maintains stable and reliable performance when dealing with different display device characteristics and diverse shooting environments, providing strong technical support for high-security scenarios such as finance. Furthermore, it possesses good scalability and adaptability, and can widely cover diverse reproduction scenarios such as various certificates and documents, and screen images.

[0127] Based on the same technical concept, this disclosure also provides a device 800 for recognizing reproduced images, such as... Figure 8 As shown, it includes: Extraction module 801 is used to extract various types of image features from the target image; The fusion module 802 is used to fuse multiple types of image features to obtain fused features; The recognition module 803 is used to classify the target image based on the fused features to obtain the re-image recognition result of the target image.

[0128] In some embodiments, the extraction module 801 includes at least two of the following units to obtain multiple types of image features: The first extraction unit 8011 is used to extract local texture features from the target image; The second extraction unit 8012 is used to extract global semantic features from the target image; The third extraction unit 8013 is used to extract frequency domain features from the target image.

[0129] In some embodiments, the first extraction unit 8011 is specifically used for: The target image is sampled to obtain intermediate images of the target image at multiple scales; Initial texture features are extracted from the intermediate images at each scale. The initial texture features corresponding to each scale are input into the corresponding convolutional neural network to obtain intermediate texture features at multiple scales. By fusing intermediate texture features at multiple scales, local texture features are obtained.

[0130] In some embodiments, the first extraction unit 8011 is specifically used for: The target image is downsampled using a first downsampling rate and a second downsampling rate to obtain an intermediate image corresponding to the first downsampling rate and an intermediate image corresponding to the second downsampling rate; the first downsampling rate is less than the second downsampling rate. The intermediate images at multiple scales also include the target image.

[0131] In some embodiments, the first extraction unit 8011 is specifically used for: An adaptive threshold is used to extract local binary pattern features from the intermediate images at each scale, which are then used as initial texture features.

[0132] In some embodiments, a first extraction unit 8011 is further included, specifically for: For the intermediate images at each scale, an adaptive threshold corresponding to the scale is determined based on the following method: For each center pixel in the intermediate image at a given scale, determine the standard deviation and mean of the pixel values ​​within a preset neighborhood of the center pixel. Based on the weight coefficients corresponding to the scale, the standard deviation, mean, and pixel value of the center pixel are weighted and summed to obtain the adaptive threshold corresponding to the scale.

[0133] In some embodiments, the sampling radius of the circular neighborhood used to extract local binary pattern features increases as the scale decreases.

[0134] In some embodiments, the first extraction unit 8011 is specifically used for: For each scale of initial texture features, perform the following operations: The encoder based on scale-corresponding convolutional neural networks encodes the initial texture features to obtain encoded features; the encoder uses various types of convolutional kernels to process the initial texture features; different types of convolutional kernels are used to extract re-image artifact features in different directions; A decoder based on a scale-corresponding convolutional neural network processes the encoded features to obtain intermediate texture features at each scale.

[0135] In some embodiments, the various types of convolution kernels include at least two of the following: Horizontal stripe detection kernel, used to detect periodic stripes in the horizontal direction; Vertical stripe detection kernel, used to detect periodic stripes in the vertical direction; Mesh detection kernel, used to detect mesh features; A diagonal cross detection kernel is used to detect periodic textures in a diagonal direction.

[0136] In some embodiments, the first extraction unit 8011 is specifically used for: The initial texture features are processed by multiple first convolutional blocks in the encoder to obtain the first intermediate features; wherein, multiple types of convolutional kernels are located in the first convolutional layer, and each type of convolutional kernel performs convolution operation on the initial texture features to construct channel features; The first intermediate feature is processed by the second convolutional block in the encoder to obtain the encoded feature.

[0137] In some embodiments, each first convolutional block includes a convolutional layer, a normalization layer, and an activation layer connected in sequence; and the output features of the previous first convolutional block are input to the next first convolutional block after pooling.

[0138] In some embodiments, the first extraction unit 8011 is specifically used for: For each scale, the encoded features are processed by a cascaded multi-layer fully connected network in the decoder corresponding to that scale to obtain the second intermediate features; The second intermediate feature is processed using a linear transform layer in the encoder to obtain the intermediate texture features at the scale.

[0139] In some embodiments, the first extraction unit 8011 is specifically used for: The attention weights are obtained by processing the spliced ​​features of local texture features at various scales based on the channel attention network. The attention weights are split into sub-weights; The feature reshaping operation is performed on each sub-weight to obtain the scale weights for each scale; Based on the weights of each scale, the intermediate texture features of each scale are weighted and fused to obtain weighted features; The weighted features are transformed to obtain local texture features.

[0140] In some embodiments, the second extraction unit 8012 is specifically used for: The target image is segmented into multiple image patches; Edge-preserving filtering is performed on each image patch to obtain the filtered features; The filtered features of each image patch are input into the convolution processing module to obtain local features; The local features of each image patch are input into the encoder to obtain global semantic features.

[0141] In some embodiments, the third extraction unit 8013 is specifically used for: The target image is processed to remove moiré patterns, resulting in an optimized image. The target image and the optimized image are transformed to the frequency domain respectively to obtain the first spectral image and the second spectral image; The difference between the first and second spectral images is determined to obtain the frequency domain features.

[0142] In some embodiments, the fusion module 802 is specifically used for: A multi-scale cross-attention fusion mechanism is used to process various types of image features to obtain fused features; Among them, local texture features are used as query features, global semantic features are used as key features, and frequency domain features are used as value features.

[0143] The specific functions and examples of each module and submodule of the apparatus in this disclosure can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0144] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0145] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0146] Figure 9 A schematic block diagram of an example electronic device 900 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0147] like Figure 9 As shown, device 900 includes a computing unit 901, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 902 or a computer program loaded from storage unit 908 into random access memory (RAM) 903. RAM 903 may also store various programs and data required for the operation of device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via bus 904. Input / output (I / O) interface 905 is also connected to bus 904.

[0148] Multiple components in device 900 are connected to I / O interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of monitors, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0149] The computing unit 901 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as the method for recognizing reproduced images. For example, in some embodiments, the method for recognizing reproduced images can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by the computing unit 901, one or more steps of the method for recognizing reproduced images described above can be performed. Alternatively, in other embodiments, the computing unit 901 can be configured to perform the method for recognizing reproduced images by any other suitable means (e.g., by means of firmware).

[0150] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0151] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0152] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0153] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0154] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0155] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0156] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0157] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for recognizing reproduced images, comprising: Extracting multiple types of image features from the target image; The various types of image features are fused to obtain fused features; Based on the fusion features, the target image is classified to obtain the re-image recognition result of the target image.

2. The method according to claim 1, wherein, The extraction of multiple types of image features from the target image includes: Perform at least two of the following operations to obtain the various types of image features: Extract local texture features from the target image; Extract global semantic features from the target image; Frequency domain features are extracted from the target image.

3. The method according to claim 2, wherein, The step of extracting local texture features from the target image includes: The target image is sampled to obtain intermediate images of the target image at multiple scales; Initial texture features are extracted from the intermediate images at each scale. The initial texture features corresponding to each scale are input into the corresponding convolutional neural network to obtain intermediate texture features at multiple scales. The local texture features are obtained by fusing the intermediate texture features at multiple scales.

4. The method according to claim 3, wherein, The step of sampling the target image to obtain intermediate images of the target image at multiple scales includes: The target image is downsampled using a first downsampling rate and a second downsampling rate to obtain an intermediate image corresponding to the first downsampling rate and an intermediate image corresponding to the second downsampling rate; the first downsampling rate is less than the second downsampling rate. The intermediate images at multiple scales also include the target image.

5. The method according to claim 3, wherein, The step of extracting initial texture features from intermediate images at each scale includes: An adaptive threshold is used to extract local binary pattern features from the intermediate images at each scale, which are then used as the initial texture features.

6. The method according to claim 5, further comprising: For each scale of intermediate image, an adaptive threshold corresponding to that scale is determined based on the following method: For each center pixel in the intermediate image at the specified scale, determine the standard deviation and mean of the pixel values ​​within a preset neighborhood of the center pixel. Based on the weighting coefficients corresponding to the scale, the standard deviation, the mean, and the pixel value of the center pixel are weighted and summed to obtain the adaptive threshold corresponding to the scale.

7. The method according to claim 5, wherein, The sampling radius of the circular neighborhood used to extract the local binary pattern features increases as the scale decreases.

8. The method according to claim 3, wherein, The step of inputting the initial texture features corresponding to each scale into the corresponding convolutional neural network to obtain intermediate texture features at multiple scales includes: For each scale of initial texture features, perform the following operations: An encoder based on the convolutional neural network corresponding to the scale encodes the initial texture features to obtain encoded features; wherein, the encoder uses multiple types of convolutional kernels to process the initial texture features; different types of convolutional kernels are used to extract re-image artifact features in different directions; The decoder of the convolutional neural network corresponding to the scale processes the encoded features to obtain the intermediate texture features of the scale.

9. The method according to claim 8, wherein, The various types of convolutional kernels include at least two of the following: Horizontal stripe detection kernel, used to detect periodic stripes in the horizontal direction; Vertical stripe detection kernel, used to detect periodic stripes in the vertical direction; Mesh detection kernel, used to detect mesh features; A diagonal cross detection kernel is used to detect periodic textures in a diagonal direction.

10. The method according to claim 8, wherein, The encoder based on the convolutional neural network corresponding to the scale encodes the initial texture features to obtain encoded features, including: The initial texture features are processed by multiple first convolutional blocks in the encoder to obtain first intermediate features; wherein, the various types of convolutional kernels are located in the first convolutional layer, and each type of convolutional kernel performs a convolution operation on the initial texture features to construct channel features; The first intermediate feature is processed using the second convolutional block in the encoder to obtain the encoded feature.

11. The method according to claim 10, wherein, Each first convolutional block includes a concatenated convolutional layer, a normalization layer, and an activation layer; and the output features of the previous first convolutional block are pooled and then input into the next first convolutional block.

12. The method according to claim 8, wherein, The decoder of the convolutional neural network corresponding to the scale processes the encoded features to obtain intermediate texture features at the scale, including: For each scale, the encoded features are processed by a cascaded multilayer fully connected network in the decoder corresponding to that scale to obtain the second intermediate features; The second intermediate feature is processed using the linear transform layer in the encoder to obtain the intermediate texture feature at the specified scale.

13. The method according to claim 3, wherein, The step of fusing the intermediate texture features at multiple scales to obtain the local texture features includes: The attention weights are obtained by processing the spliced ​​features of local texture features at various scales based on the channel attention network. The attention weights are split to obtain sub-weights; The feature reshaping operation is performed on each sub-weight to obtain the scale weights for each scale; Based on the weights of each scale, the intermediate texture features of each scale are weighted and fused to obtain weighted features; The local texture features are obtained by performing a feature transformation operation on the weighted features.

14. The method according to claim 2, wherein, Extracting global semantic features from the target image includes: The target image is divided into multiple image blocks; Edge-preserving filtering is performed on each image patch to obtain the filtered features; The filtered features of each image patch are input into the convolution processing module to obtain local features; The local features of each image patch are input into the encoder to obtain the global semantic features.

15. The method according to claim 2, wherein, Extracting frequency domain features from the target image includes: The target image is subjected to moiré removal processing to obtain an optimized image; The target image and the optimized image are transformed to the frequency domain respectively to obtain a first spectral image and a second spectral image; The difference between the first spectral image and the second spectral image is determined to obtain the frequency domain features.

16. The method according to claim 2, wherein, The process of fusing the various types of image features to obtain fused features includes: The multi-scale cross-attention fusion mechanism is used to process the various types of image features to obtain the fused features; The local texture features serve as query features, the global semantic features serve as key features, and the frequency domain features serve as value features.

17. A device for recognizing reproduced images, comprising: The extraction module is used to extract various types of image features from the target image; The fusion module is used to perform fusion operations on the various types of image features to obtain fused features; The recognition module is used to perform a classification operation on the target image based on the fused features to obtain the re-image recognition result of the target image.

18. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-16.

19. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-16.

20. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-16.

Images