Image propagation chain detection method and device based on dynamic information compensation and computer

Abstract

The invention discloses an image propagation chain detection method and device based on dynamic information compensation and a computer, and the method comprises the steps: obtaining a to-be-processed image, and carrying out the feature extraction of the image, and obtaining a noise domain feature and a frequency domain feature; fusing the noise domain features and the frequency domain features, and inputting the fused features into an encoder to extract composite features of the image; and extracting compensation features from a preset dynamic information compensation mechanism, combining the compensation features with the composite features to realize preorder platform trace enhancement so as to improve the detection precision of the image during platform propagation, and inputting the combined composite features into a decoder to detect the next platform. According to the method, a memory bank is established through a dynamic information compensation mechanism, and part of preorder platform traces lost by subsequent platform processing in the propagation process of the image are adaptively and dynamically learned, so that the detection precision of the image during platform propagation is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to an image propagation chain detection method, device, and computer device based on dynamic information compensation. Background Art

[0002] With the popularity of social platforms (such as WeChat, Facebook, WhatsApp, etc.), the propagation of images among multiple platforms has become a common phenomenon. Each platform usually performs a series of customized processing operations on the uploaded images, such as renaming, recompressing, resizing, etc. These operations will leave unique "traces" in the images. These traces provide important clues for image propagation chain detection, enabling the complete reconstruction of the propagation path of the image on the social platform in chronological order. Image propagation chain detection is an important research task in the field of multimedia forensics. In the current network environment, the illegal propagation of images brings serious security threats, including issues such as copyright infringement, privacy leakage, and reputation damage. Therefore, monitoring the entire process of image propagation is crucial for maintaining network security.

[0003] When studying the problem of the image propagation chain, the following main difficulties exist: 1) Concealment: During the propagation process of the image, the traces left by the processing of the previous platform are often covered or weakened by the operations of the subsequent platform, such as compression and resizing, resulting in the unobvious traces of platform processing. 2) Incompleteness: During the propagation process of the image, it usually undergoes compression operations by the platform. However, existing algorithms usually only rely on the DCT histogram to extract compression information. On the one hand, it is difficult to comprehensively perceive the compression features; on the other hand, due to the lack of effective utilization of spatial information, it is difficult to capture the correlation characteristics of pixel-level modifications and compression changes between pixel blocks during the compression process. Summary of the Invention

[0004] The present invention provides an image propagation chain detection method based on dynamic information compensation, including:

[0005] Obtain the image to be processed, and respectively perform feature extraction on the image to obtain noise domain features and frequency domain features;

[0006] Fuse the noise domain features and the frequency domain features and input them into an encoder to extract the composite features of the image;

[0007] Extract compensation features from a preset dynamic information compensation mechanism, and combine the compensation features with the composite features to enhance the traces of the previous platform to improve the detection accuracy of the image during platform propagation, and input the combined composite features into a decoder to detect the next platform.

[0008] Further, the respectively performing feature extraction on the image to obtain noise domain features and frequency domain features includes:

[0009] Extract the noise information of the image through the SRM conversion model, and extract the features of the noise information through the Backbone_N model to generate the noise domain features;

[0010] Extract the binary stereo DCT information of the image through the DCT conversion model, and extract the features through the Backbone_D model to generate the frequency domain features.

[0011] Further, the fusing the noise domain features and the frequency domain features and inputting the fused features into an encoder to extract the composite features of the image includes:

[0012] Concatenate the noise domain features and the frequency domain features by channel and input them into a fusion model for feature fusion;

[0013] Concatenate the fused features with the container features and input them into a Transformer encoder to obtain the composite features of the image.

[0014] Further, the dynamic information compensation mechanism includes: a dynamic memory bank for storing context and compensation features and an addressing module for accessing the dynamic memory bank, the compensation features being used to capture feature information lost during processing during the propagation of the image on a previous platform, and the context representing the encoded information of the propagation chain.

[0015] Further, extracting the compensation features from a preset dynamic information compensation mechanism and combining the compensation features with the composite features includes:

[0016] The addressing module uses the propagation sub-chain input by the decoder as a query condition to retrieve the most relevant context information from the storage unit of the dynamic memory bank and extract the corresponding compensation features;

[0017] Combine the compensation features with the composite feature set to enhance the composite features.

[0018] Further, extracting the compensation features from a preset dynamic information compensation mechanism and combining the compensation features with the composite features further includes:

[0019] Adopt the gradient descent method to minimize the cross-entropy loss function and continuously optimize and update the dynamic information compensation mechanism.

[0020] The present invention provides an image propagation chain detection device based on dynamic information compensation, including:

[0021] A feature acquisition module, configured to acquire an image to be processed, and respectively perform feature extraction on the image to obtain noise domain features and frequency domain features;

[0022] A feature processing module, configured to fuse the noise domain features and the frequency domain features and input the fused features into an encoder to extract composite features of the image;

[0023] A feature detection module, configured to extract compensation features from a preset dynamic information compensation mechanism, combine the compensation features with the composite features to enhance the previous platform traces and improve the detection accuracy of the image during platform propagation, and input the combined composite features into a decoder to detect the next platform.

[0024] Further, the feature acquisition module includes:

[0025] A first feature acquisition sub-module, configured to extract noise information of the image through an SRM conversion model, and perform feature extraction on the noise information through a Backbone_N model to generate the noise domain features;

[0026] A second feature acquisition sub-module, configured to extract binary stereo DCT information of the image through a DCT conversion model, and perform feature extraction through a Backbone_D model to generate the frequency domain features.

[0027] Further, the feature processing module includes:

[0028] A first feature processing sub-module, configured to splice the noise domain features and the frequency domain features by channel and input the spliced features into a fusion model for feature fusion;

[0029] A second feature processing sub-module, configured to splice the fused features with container features and input the spliced features into a Transformer encoder to obtain the composite features of the image.

[0030] Further, the dynamic information compensation mechanism includes: a dynamic memory bank for storing context and compensation features, and an addressing module for accessing the dynamic memory bank. The compensation features are used to capture feature information lost during processing of the image during propagation on the previous platform, and the context represents the encoded information of the propagation chain.

[0031] Further, the feature detection module includes:

[0032] A first feature detection sub-module, configured to use the propagation sub-chain input to the decoder as a query condition, retrieve the most relevant context information from the storage unit of the dynamic memory bank, and extract the corresponding compensation features;

[0033] A second feature detection sub-module, configured to combine the compensation features with the composite feature set to enhance the composite features.

[0034] Further, the feature detection module further includes:

[0035] The third feature detection sub-module is used to continuously optimize and update the dynamic information compensation mechanism by minimizing the cross-entropy loss function using the gradient descent method.

[0036] The present invention provides a computer device, including a memory and a processor. Computer-readable instructions are stored in the memory. When the computer-readable instructions are executed by the processor, the processor executes the steps of the image propagation chain detection method based on dynamic information compensation as described above.

[0037] The present invention provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the image propagation chain detection method based on dynamic information compensation as described above.

[0038] The beneficial effects of the embodiments of the present invention are as follows: The present invention establishes a memory bank through a dynamic information compensation mechanism, and adaptively and dynamically learns the part of the pre-order platform traces lost during the image propagation process due to subsequent platform processing. In the inference stage, the stored compensation information is extracted from the memory bank to enhance the image features, thereby effectively strengthening the pre-order platform traces in the image. To solve the incompleteness problem, the present invention further proposes a two-stream image feature extraction framework, which comprehensively captures compression features by simultaneously extracting the compressed information of the image in the frequency domain and the residual domain. In addition, a binary stereo DCT is used to replace the traditional DCT histogram, focusing on the pixel-level modifications generated during the compression process, enhancing the perception ability of the spatial changes caused by the compression operation, and improving the extraction effect of the platform compression features. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative efforts.

[0040] Figure 1 It is a flowchart of an image propagation chain detection method based on dynamic information compensation provided by the present invention;

[0041] Figure 2 It is a schematic diagram of the network framework of an image propagation chain detection method based on dynamic information compensation provided by the present invention;

[0042] Figure 3 It is a schematic diagram of the Backbone_D model structure provided by the present invention;

[0043] Figure 4Schematic diagram of the Backbone_N model structure provided by the present invention;

[0044] Figure 5 Schematic diagram of the SRM structure provided by the present invention;

[0045] Figure 6 Schematic diagram of the ResBlock structure provided by the present invention;

[0046] Figure 7 Basic structural block diagram of an image propagation chain detection device based on dynamic information compensation provided by the present invention;

[0047] Figure 8 Basic structural block diagram of the computer device provided by the present invention. Detailed implementation manners

[0048] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0049] As Figure 1 shown, the present invention provides an image propagation chain detection method based on dynamic information compensation, including:

[0050] S1. Obtain the image to be processed, and respectively perform feature extraction on the image to obtain the noise domain feature and the frequency domain feature;

[0051] S2. After fusing the noise domain feature and the frequency domain feature, input them into an encoder to extract the composite feature of the image;

[0052] S3. Extract the compensation feature from a preset dynamic information compensation mechanism, and combine the compensation feature with the composite feature to enhance the traces of the previous platform to improve the detection accuracy when the image is propagated on the platform, and input the combined composite feature into a decoder to detect the next platform.

[0053] In step S1 of the present invention, respectively performing feature extraction on the image to obtain the noise domain feature and the frequency domain feature includes:

[0054] Step 1. Extract the noise information of the image through the SRM conversion model, and perform feature extraction on the noise information through the Backbone_N model to generate the noise domain feature;

[0055] Step 2. Extract the binary stereo DCT information of the image through the DCT conversion model, and perform feature extraction through the Backbone_D model to generate the frequency domain feature.

[0056] Please refer to Figure 2As shown, the present invention extracts the noise information and DCT information of the input image through the SRM and DCT conversion models respectively. The DCT conversion model also converts the DCT information into binary stereo DCT information to comprehensively capture the compression features during the image propagation process. Then, these features are respectively subjected to feature extraction through Backbone_N and Backbone_D to generate noise domain features and frequency domain features. Specifically, the process of dual-stream image feature extraction is as follows:

[0057] 1. Extraction of frequency domain stream features:

[0058] The image to be processed is converted into the DCT representation D through the DCT conversion model. Subsequently, D is converted into the binary stereo DCT form D'. The specific operation is as follows: First, D ∈ R 256×256 is cropped, and the range is limited to the threshold [-T, T], where T is 20. Then, the cropped values are binarized and transformed into the {0, 1} 256×256×(T+1) form. The specific formula is as follows:

[0059]

[0060] where t ∈ [0, T], |·| represents the element-wise absolute value, and clip(·) represents the operation of constraining D within the range [-T, T]: clip(D i,j ) = max(-T, min(D i,j , T)).

[0061] Subsequently, D' and the quantization table are passed through Backbone_D, as Figure 3 shown, to obtain features of 32×32×128. Backbone_D first passes through a dilated convolution with a dilation rate of 8 and a 1×1 convolution to obtain the feature D 1 , then copies the quantization table Q to the dimension 256×256×1, and multiplies it with D 1 to obtain the feature D 2 . Then, D 1 and D 2 are respectively reshaped into features of 32×32×256 and concatenated. Finally, the concatenated features are processed through a common convolution layer to obtain the frequency domain feature D 0 .

[0062] 2. Extraction of noise domain stream features:

[0063] First, the three channels of the image to be processed, as Figure 5 shown, are respectively processed through the three convolution layers in the SRM model to generate 9 feature maps. Subsequently, these feature maps are concatenated by channel and input into Backbone_N, asFigure 4 As shown, first, it goes through 4 convolutional layers to obtain N 1 , then it is input into the ResBlock and the max pooling layer to obtain N 2 , and continues to pass through the ResBlock to obtain N 3 , and finally goes through another max pooling layer to obtain N 4 , where the dimensions of N 1 , N 2 , N 3 and N 4 are all 32×32×64. Then, they are concatenated by channel and passed through two convolutional layers, and then through the cSElayer, and finally the noise feature S 0 is obtained, and its dimension is 32×32×64. Among them, the ResBlock is as shown in Figure 6 .

[0064] In the S2 embodiment of the present invention, after fusing the noise domain feature and the frequency domain feature, they are input into the encoder to extract the composite feature of the image, including:

[0065] Step 1: Concatenate the noise domain feature and the frequency domain feature by channel and input them into the fusion model for feature fusion;

[0066] Step 2: Concatenate the fused feature with the container feature and input it into the Transformer encoder to obtain the composite feature of the image.

[0067] In one embodiment of the present invention, the frequency domain feature D 0 and the noise feature S 0 are concatenated by channel and then input into the fusion module. The fusion module is composed of fully connected layers, which maps the feature dimension from 32×32×(64 + 128) to 32×32×64, and finally flattens it to 1024×64. Then, it is concatenated with the container feature of the image to obtain F 0 , and finally it is input into the encoder to obtain the composite feature z∈R 512×64 . It should be noted that the Transformer encoder selects a Transformer encoder without attention.

[0068] In the embodiment of the present invention, the dynamic information compensation mechanism includes: a dynamic memory bank for storing context and compensation features and an addressing module for accessing the dynamic memory bank. The compensation feature is used to capture the feature information lost due to processing during the propagation of the image on the previous platform, and the context represents the encoded information of the propagation chain.

[0069] Extracting the compensation feature from the preset dynamic information compensation mechanism and combining the compensation feature with the composite feature, including:

[0070] Step 1: The addressing module uses the propagator chain input by the decoder as a query condition, retrieves the most relevant context information from the storage units of the memory dynamic memory bank, and extracts the corresponding compensation features;

[0071] Step 2: Combine the compensation features with the composite feature set to enhance the composite features.

[0072] An embodiment of the present invention, as Figure 2 shown, the dynamic information compensation mechanism includes: a dynamic memory bank and an addressing module.

[0073] The dynamic memory bank consists of M storage units, and each storage unit contains two types of information: compensation feature c m and context a m , where the compensation feature is used to capture the feature information lost during the propagation of the image on the previous platform, and the context represents the encoded information of the propagator chain.

[0074] Assume there are L types of propagation chains, and the i-th propagator chain of the l-th propagation chain is assigned to the m-th index of the storage unit, where 1 ≤ m ≤ M. The i-th propagator chain is shown as follows:

[0075]

[0076] where n represents the number of propagation times of the l-th propagation chain, and O j represents the j-th platform on the propagator chain In this embodiment, a is randomly initialized using the embedding vector of m , and c m is initialized using a zero vector.

[0077] The addressing module uses the propagator chain input by the decoder as a query, retrieves the most relevant context information from the storage units of the memory bank, and extracts the corresponding compensation features.

[0078] For the image to be processed, the encoder generates a composite feature z l . And at time t, the propagator chain of the corresponding propagation chain is decoded

[0079]

[0080] where, represents the j-th platform on the decoded propagator chain .

[0081] Use the compensation features in the memory bank to enhance z l . Specifically, for the decoding process at time t, the composite feature Z land the decoded propagator chain is input into the following equation to generate a query encoding

[0082]

[0083] where z l = encoder(F 0 ), mean(·) represents taking the average in the first dimension, PE(·) represents positional encoding, W q , W K and W V represent weight matrices, T represents the transpose operation, and d k represents the dimension of K.

[0084] Subsequently, the top K context-corresponding compensation features most similar to the query encoding are selected from the memory bank to enhance z l , and their weights are determined using the Softmax function:

[0085]

[0086] where d(·,·) represents cosine similarity, and a k represents the contexts of the top K storage units with the highest similarity. Then, the compensation feature of the decoded propagator chain at time t is calculated

[0087]

[0088] where c k is the compensation feature corresponding to a k .

[0089] The is weighted and summed with z l to obtain an enhanced composite feature The memory bank adaptively learns c m and a m during the training process.

[0090] The composite feature and the decoded propagator chain are input into a non-attentive Transformer decoder for predicting the next platform. An embodiment of the present invention further includes:

[0091] Adopting the gradient descent method to minimize the cross-entropy loss function to continuously optimize and update the dynamic information compensation mechanism.

[0092] where the loss function is Optimize the neural network through the cross-entropy loss function, making the predicted propagation chain approximate the true propagation chain y.

[0093] As shown in Table 1, the R-SMUD and V-SMUD datasets used are existing image propagation chain datasets. Each column shows the accuracy of evaluating the model on subsets of the two datasets with propagation chain lengths from 1 to 3. For example, C3 means the picture is propagated across 3 platforms. The metric used in the table is accuracy, which is used to measure the accuracy of the model's detection results.

[0094] Table 1. Comparison of Experimental Results of Different Image Propagation Chain Detection Methods (Accuracy %)

[0095]

[0096] Table 1 compares the experimental results of different image propagation chain detection methods. Based on these results, the following conclusions are drawn: The method proposed in the present invention (Ours) is superior to the current state-of-the-art methods on all propagation sub-chains of each dataset. Compared with the P-CNN-FF algorithm of the convolutional neural network and the Ve method based on the random forest classifier, the method proposed in the present invention has achieved a significant improvement in detection performance. For example, for the C3 subset of the R-SMUD and V-SMUD datasets, neither the P-CNN-FF nor the Ve method could exceed an accuracy of 50%, while the detection accuracies of the method proposed in the present invention were 60.35% and 65.42% respectively. This result shows that the two-stream image propagation chain detection method based on the dynamic information compensation mechanism designed in the present invention can better capture the correlation between platforms compared with traditional classification methods, thus improving the detection performance. Compared with the sequence-to-sequence image propagation chain detection algorithm Seq2Seq, the method proposed in the present invention has obtained a large performance improvement. For example, the accuracies of Seq2Seq on the C3 subsets of R-SMUD and V-SMUD were 50.68% and 59.68% respectively, while the prediction accuracies of the present invention were 60.35% and 65.42% respectively, an increase of 9.67% and 5.74% respectively, indicating that the two-stream image propagation chain detection method based on the dynamic information compensation mechanism designed in the present invention can better enhance the trace loss of the previous platform in the image propagation process compared with Seq2Seq, thus improving the detection performance.

[0097] Therefore, the present invention proposes and designs a two-stream image feature extraction method to simultaneously extract the compression traces of the image from the frequency domain and the residual domain, realizing the comprehensive capture of the compression features. In the frequency domain, binary stereo DCT is used to replace the traditional DCT histogram, which focuses on the pixel-level modifications caused by the compression operation and is beneficial to perceiving the spatial changes in compression.

[0098] In addition, the dynamic information compensation mechanism proposed by the present invention uses the decoded propagator chain information to dynamically enhance the traces of the previous platform, thereby improving the accuracy of propagation chain detection.

[0099] As Figure 7 shown, the present invention provides an image propagation chain detection device based on dynamic information compensation, including: a feature acquisition module 2100, configured to acquire an image to be processed, and respectively perform feature extraction on the image to obtain a noise domain feature and a frequency domain feature; a feature processing module 2200, configured to fuse the noise domain feature and the frequency domain feature and input them into an encoder to extract a composite feature of the image; a feature detection module 2300, configured to extract a compensation feature from a preset dynamic information compensation mechanism, and combine the compensation feature with the composite feature to enhance the traces of the previous platform to improve the detection accuracy when the image is propagated on the platform, and input the combined composite feature into a decoder to detect the next platform.

[0100] In some embodiments, the feature acquisition module 2100 includes: a first feature acquisition sub-module, configured to extract the noise information of the image through an SRM conversion model, and perform feature extraction on the noise information through a Backbone_N model to generate the noise domain feature; a second feature acquisition sub-module, configured to extract the binary stereo DCT information of the image through a DCT conversion model, and perform feature extraction through a Backbone_D model to generate the frequency domain feature.

[0101] In some embodiments, the feature processing module 2200 includes: a first feature processing sub-module, configured to splice the noise domain feature and the frequency domain feature according to channels and input them into a fusion model for feature fusion; a second feature processing sub-module, configured to splice the fused feature with a container feature and input it into a Transformer encoder to obtain a composite feature of the image.

[0102] In some embodiments, the dynamic information compensation mechanism includes: a dynamic memory bank for storing context and compensation features, and an addressing module for accessing the dynamic memory bank, where the compensation feature is used to capture the feature information lost during the processing of the image during the propagation process on the previous platform, and the context represents the encoded information of the propagation chain.

[0103] In some embodiments, the feature detection module 2300 includes: a first feature detection sub-module, configured to use the propagator chain input to the decoder as a query condition to retrieve the most relevant context information from the storage unit of the dynamic memory bank, and extract the corresponding compensation feature; a second feature detection sub-module, configured to combine the compensation feature with the composite feature set to enhance the composite feature.

[0104] In some embodiments, the feature detection module 2300 further includes a third feature detection sub-module, which is configured to continuously optimize and update the dynamic information compensation mechanism by minimizing the cross-entropy loss function using the gradient descent method.

[0105] The present invention establishes a memory bank through a dynamic information compensation mechanism, and adaptively and dynamically learns the traces of the previous platforms that are lost due to subsequent platform processing during the propagation of images. In the inference stage, the stored compensation information is extracted from the memory bank to enhance the image features, thereby effectively strengthening the traces of the previous platforms in the image. To solve the problem of incompleteness, the present invention further proposes a dual-stream image feature extraction framework, which comprehensively captures the compression features by simultaneously extracting the compressed information of the image in the frequency domain and the residual domain. In addition, a binary stereo DCT is used to replace the traditional DCT histogram, focusing on the pixel-level modifications generated during the compression process, enhancing the perception ability of the spatial changes caused by the compression operation, and improving the extraction effect of the platform compression features.

[0106] To solve the above technical problems, an embodiment of the present invention also provides a computer device. Specifically, please refer to Figure 8 , Figure 8 which is the basic structural block diagram of the computer device in this embodiment.

[0107] As Figure 8 shown, it is the internal structural schematic diagram of the computer device. As Figure 8 shown, the computer device includes a processor, a non-volatile storage medium, a memory, and a network interface connected through a system bus. Among them, the non-volatile storage medium of the computer device stores an operating system, a database, and computer-readable instructions. The control information sequence can be stored in the database. When the computer-readable instructions are executed by the processor, the processor can implement an image propagation chain detection method based on dynamic information compensation. The processor of the computer device is used to provide computing and control capabilities to support the operation of the entire computer device. The memory of the computer device can store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor can execute an image processing method. The network interface of the computer device is used to connect and communicate with the terminal. Those skilled in the art can understand that Figure 8 the structure shown in

[0108] In this embodiment, the processor is used to execute Figure 7For the specific content of the feature acquisition module 2100, the feature processing module 2200, and the feature detection module 2300, the memory stores the program codes and various types of data required to execute the above modules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program codes and data required to execute all sub-modules in the image propagation chain detection method based on dynamic information compensation, and the server can call the program codes and data of the server to execute the functions of all sub-modules.

[0109] The computer device provided by the embodiment of the present invention establishes a memory bank through a dynamic information compensation mechanism, and adaptively and dynamically learns the part of the pre-order platform traces lost during the propagation process of the image due to subsequent platform processing. In the inference stage, the stored compensation information is extracted from the memory bank to enhance the image features, thereby effectively strengthening the pre-order platform traces in the image. To solve the incompleteness problem, the present invention further proposes a two-stream image feature extraction framework, which comprehensively captures the compressed features by simultaneously extracting the compressed information of the image in the frequency domain and the residual domain. In addition, binary stereo DCT is used to replace the traditional DCT histogram, focusing on the pixel-level modifications generated during the compression process, enhancing the perception ability of the spatial changes caused by the compression operation, and improving the extraction effect of the platform compression features.

[0110] The present invention also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors are caused to execute the steps of the image propagation chain detection method based on dynamic information compensation according to any one of the above embodiments.

[0111] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0112] It should be understood that although the steps in the flowchart of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction for the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0113] The above are only some embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. An image propagation chain detection method based on dynamic information compensation, characterized in that: include: Acquire an image to be processed, and extract features of the image to obtain noise domain features and frequency domain features; The noise domain features and the frequency domain features are fused and input into an encoder to extract the composite features of the image; A compensation feature is extracted from a preset dynamic information compensation mechanism, and the compensation feature is combined with the composite feature to enhance the preceding platform traces to improve the detection accuracy of the image when it is propagated on the platform, and the combined composite feature is input into the decoder to detect the next platform.

2. The detection method according to claim 1, characterized in that: The extracting features of the image to obtain noise domain features and frequency domain features respectively includes: Extracting noise information of the image through the SRM conversion model, and performing feature extraction on the noise information through the Backbone_N model to generate the noise domain feature; The binary stereo DCT information of the image is extracted through the DCT conversion model, and the frequency domain features are generated by performing feature extraction through the Backbone_D model.

3. The detection method according to claim 1, characterized in that: The step of fusing the noise domain features and the frequency domain features and inputting the fused features into an encoder to extract the composite features of the image comprises: Inputting the noise domain features and the frequency domain features into a fusion model according to channel splicing for feature fusion; The fused features are concatenated with the container features and input into the Transformer encoder to obtain the composite features of the image.

4. The detection method according to claim 1, characterized in that: The dynamic information compensation mechanism includes: a dynamic memory library for storing context and compensation features and an addressing module for accessing the dynamic memory library, the compensation features are used to capture the feature information of the image lost due to processing during the propagation process of the previous platform, and the context represents the encoding information of the propagation chain.

5. The detection method according to claim 4, characterized in that: The extracting the compensation feature from the preset dynamic information compensation mechanism and combining the compensation feature with the composite feature includes: The addressing module uses the propagation subchain input by the decoder as a query condition, retrieves the most relevant context information from the storage unit of the dynamic memory bank, and extracts the corresponding compensation features; The compensating feature is combined with the composite feature to enhance the composite feature.

6. The detection method according to claim 4, characterized in that: Extracting a compensation feature from a preset dynamic information compensation mechanism and combining the compensation feature with the composite feature, further comprising: The gradient descent method is used to minimize the cross entropy loss function, and the dynamic information compensation mechanism is continuously optimized and updated.

7. An image propagation chain detection device based on dynamic information compensation, characterized in that: include: A feature acquisition module is used to acquire an image to be processed, and extract features from the image to obtain noise domain features and frequency domain features; A feature processing module, used for fusing the noise domain features and the frequency domain features and inputting them into an encoder to extract the composite features of the image; The feature detection module is used to extract compensation features from a preset dynamic information compensation mechanism, and combine the compensation features with the composite features to enhance the previous platform traces to improve the detection accuracy of the image when it is propagated on the platform, and input the combined composite features into the decoder to detect the next platform.

8. The device according to claim 7, characterized in that The feature acquisition module comprises: A first feature acquisition submodule is used to extract noise information of the image through an SRM conversion model, and to perform feature extraction on the noise information through a Backbone_N model to generate the noise domain feature; The second feature acquisition submodule is used to extract the binary stereo DCT information of the image through the DCT conversion model, and to perform feature extraction through the Backbone_D model to generate the frequency domain feature.

9. A computer device comprising a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor performs the steps of the image propagation chain detection method based on dynamic information compensation as described in any one of claims 1 to 6.

10. A storage medium storing computer-readable instructions, which, when executed by one or more processors, enables the one or more processors to perform the steps of the image propagation chain detection method based on dynamic information compensation as described in any one of claims 1 to 6.

Images