Data processing methods, training methods for image processing systems, and electronic equipment

Abstract

This application discloses a data processing method, an image processing system training method, and an electronic device in the field of artificial intelligence, which can reduce the processing complexity of image source identification and improve processing efficiency. The data processing method includes: acquiring first acquisition trace information of an image to be processed, wherein the first acquisition trace information is used to indicate the acquisition traces formed in the image to be processed by the target acquisition device when acquiring the image to be processed; and classifying the first acquisition trace information using a classification module to determine the target acquisition device information for acquiring the image to be processed, wherein the classification module is a trained neural network model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, specifically to a data processing method, a training method for an image processing system, and an electronic device. Background Technology

[0002] Due to manufacturing variations in image acquisition equipment, traces of the equipment will remain in the acquired images. For example, noise in the image can be caused by photoresponse nonuniformity (PRNU) due to differences in the thickness of silicon on the photosensitive element, and distortion and aberrations can be caused by lens manufacturing variations.

[0003] Image source identification (ISI) technology identifies the acquisition device information of an image by recognizing acquisition traces within it. Specifically, denoising algorithms can be used to determine acquisition trace information in the image to be processed. These algorithms can also be used to determine the acquisition trace information of a particular image acquisition device based on multiple images acquired by that device. By comparing the acquisition trace information in the image to be processed with the device acquisition trace information, it can be determined whether that device was the one that acquired the image.

[0004] To determine the acquisition device of a particular image among multiple acquisition devices, it is necessary to compare the acquisition trace information of the image with the device acquisition trace information of the multiple acquisition devices, which is a relatively complex process. Summary of the Invention

[0005] This application provides a training method, data processing method, and apparatus for an image processing system, which can reduce the processing complexity of image source recognition and improve processing efficiency.

[0006] This application provides a data processing method, including:

[0007] Acquire first acquisition trace information of the image to be processed, wherein the first acquisition trace information is used to indicate the acquisition trace formed in the image to be processed by the target acquisition device when acquiring the image to be processed;

[0008] The first acquisition trace information is classified using a classification module to determine the target acquisition device information for acquiring the image to be processed. The classification module is a trained neural network model.

[0009] Optionally, obtaining the first acquisition trace information of the image to be processed includes: extracting the first acquisition trace information of the image to be processed using a trace extraction module, wherein the trace extraction module is a trained neural network model;

[0010] The trace extraction module and the classification module belong to the image processing system, which is trained using training data, including training images and information about the marker acquisition devices corresponding to the training images.

[0011] Optionally, the image processing system is obtained by training an initial image processing system using the training data. The initial image processing system includes an initial trace extraction module, which is obtained by pre-training using pre-training data. The pre-training data includes multiple pre-training images and marker acquisition trace information corresponding to each pre-training image. The trace extraction module is obtained based on the initial trace extraction module.

[0012] Optionally, the first acquisition trace information includes the second acquisition trace information of each of the plurality of image blocks in the image to be processed, wherein the plurality of image blocks have different sizes.

[0013] Optionally, obtaining the first acquisition trace information of the image to be processed includes: extracting the second acquisition trace information of each image block using the trace extraction module.

[0014] Optionally, the step of classifying the first acquisition trace information using the classification module to determine the target acquisition device information for acquiring the image to be processed includes: performing the following processing using the classification module:

[0015] Feature extraction processing is performed on the second acquisition trace information of each image block to obtain the feature information of the second acquisition trace information of each image block;

[0016] All the obtained feature information is fused to obtain fused features;

[0017] The fused features are classified to obtain information about the target acquisition device.

[0018] Optionally, the classification module includes multiple layers and a fusion submodule;

[0019] The first layer of the plurality of layers is used to extract features from the second acquisition trace information to obtain the feature extraction result of the first layer;

[0020] The layers other than the first layer are used to extract features from the feature extraction results determined by the previous layer, so as to obtain the feature extraction results of the other layers.

[0021] The fusion submodule is used to fuse the feature extraction results of multiple layers to obtain the feature information of the second acquisition trace information.

[0022] Optionally, the target acquisition device information is used to indicate the target acquisition device for acquiring the image to be processed.

[0023] This application also provides a training method for an image processing system, including:

[0024] Acquire training acquisition trace information of the training images and label acquisition device information of the training images;

[0025] The training acquisition trace information is classified using an initial classification module to obtain training acquisition device information. The initial classification module is a neural network model.

[0026] Based on the difference between the training acquisition device information and the labeled acquisition device information, the parameters of the initial classification module are adjusted to obtain a classification module. The classification module is used to classify the first acquisition trace information of the image to be processed in order to determine the target acquisition device information for acquiring the image to be processed.

[0027] The image processing system is obtained according to the classification module, and the image processing system is used to output target acquisition device information for acquiring the image to be processed based on the input image to be processed.

[0028] Optionally, obtaining the training acquisition trace information of the training image includes: extracting the training acquisition trace information using an initial trace extraction module, wherein the initial trace extraction module is a neural network model;

[0029] The step of adjusting the parameters of the initial classification module based on the difference between the training acquisition device information and the label acquisition device information includes: adjusting the parameters of the initial trace extraction module and the initial classification module based on the difference between the training acquisition device information and the label acquisition device information to obtain the trace extraction module and the classification module;

[0030] The step of obtaining the image processing system based on the classification module includes: obtaining the image processing system based on the trace extraction module and the classification module.

[0031] Optionally, the initial trace extraction module is obtained by pre-training using pre-training data, which includes multiple pre-training images and the marker acquisition trace information corresponding to each pre-training image.

[0032] Optionally, the step of extracting the training acquisition trace information using the initial trace extraction module includes: extracting the training image block acquisition trace information of each image block in a plurality of image blocks of the training image using the initial trace extraction module, wherein the training acquisition trace information includes the training image block acquisition trace information of each image block;

[0033] The step of classifying the training acquisition trace information using the initial classification module includes: performing the following processing using the initial classification module:

[0034] Feature extraction processing is performed on the training image block acquisition trace information of each image block to obtain the training feature information of the training image block acquisition trace information of each image block;

[0035] All the obtained training feature information is fused to obtain training fusion features;

[0036] The training fusion features are classified to obtain the training acquisition device information.

[0037] Optionally, the marker acquisition device information is used to indicate the training acquisition device that acquired the training images.

[0038] This application also provides an electronic device, including:

[0039] Processor; and

[0040] A memory for storing program instructions, which the image processing device executes via the processor to perform the above-described method.

[0041] This application also provides an image source tracing method, including:

[0042] Obtain the image to be traced sent by the client.

[0043] Extract the first acquisition trace information of the image to be traced, the first acquisition trace information being used to indicate the acquisition traces formed in the image to be traced by the target acquisition device when acquiring the image to be traced;

[0044] The first acquisition trace information is classified using a classification module to determine the target acquisition device information for acquiring the image to be traced.

[0045] The source tracing result is sent to the client, and the source tracing result includes the target acquisition device information.

[0046] This application also provides an image identification method, including:

[0047] Acquire the image to be identified and the information of the acquisition device to be determined;

[0048] Extract the first acquisition trace information of the image to be identified, wherein the first acquisition trace information is used to indicate the acquisition trace formed in the image to be identified by the target acquisition device when acquiring the image to be identified;

[0049] The first acquisition trace information is classified using a classification module to determine the target acquisition device information for acquiring the image to be identified.

[0050] A judgment result is sent to the client, which indicates whether the target acquisition device information is consistent with the pending acquisition device information.

[0051] According to the data processing scheme provided in the embodiments of the present invention, the target acquisition device information for acquiring the image to be processed is determined by classifying the first acquisition trace information of the image to be processed using a classification module. Thus, the determination of the target acquisition device information does not depend on comparing the first acquisition trace information with the device acquisition trace information of the acquisition device, thereby reducing the processing complexity of image source identification and improving processing efficiency. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of image acquisition and individual recognition;

[0053] Figure 2 This is a schematic flowchart of an image processing method provided in an embodiment of this application;

[0054] Figure 3 This is a schematic structural diagram of an image processing system provided in an embodiment of this application;

[0055] Figure 4 This is a schematic structural diagram of a feature extraction network provided in an embodiment of this application;

[0056] Figure 5 This is a schematic structural diagram of another image processing system provided in the embodiments of this application;

[0057] Figure 6 This is a schematic structural diagram of a trace extraction network provided in an embodiment of this application.

[0058] Figure 7 This is a schematic structural diagram of another trace extraction network provided in an embodiment of this application;

[0059] Figure 8 This is a schematic flowchart of a training method for an image processing system provided in an embodiment of this application;

[0060] Figure 9 This is a schematic flowchart of a training method for an image processing system provided in an embodiment of this application;

[0061] Figure 10 This is a schematic structural diagram of an electronic device provided in an embodiment of this application;

[0062] Figure 11This is a schematic structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0063] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0064] With the development of digital technology, images can be edited using various image editing tools, making image creation increasingly common and easy. However, at the same time, the authenticity of images has been greatly challenged.

[0065] Digital image forensics, through the analysis of statistical features of images, can determine the authenticity, integrity, and originality of image content. Digital image forensics includes image source identification (ISI) and image forgery detection. Image forgery detection technology is used to determine whether a digital image has been modified. ISI technology identifies acquisition traces in an image, revealing information about the image acquisition device. Digital image forensics has a wide range of applications.

[0066] Some images containing trade secrets or personal privacy may have been illegally collected. In cases of such image leaks, tracing the source by identifying the device used to collect the image can provide clues to determine who is responsible for the leak.

[0067] Image theft is the act of illegally using another person's image for personal or commercial purposes without the original author's permission or consent. Identifying the device used to capture the image can help determine if image theft has occurred.

[0068] Computer-generated images are generally not permitted in official documents. However, in judicial proceedings, natural images can be used as evidence. Techniques that distinguish between natural and computer-generated images can be used to determine whether an image is natural, i.e., whether it possesses authenticity, thereby determining whether the image can be used in official documents or as evidence.

[0069] Image source identification (ISI) technology has always been a research hotspot in the field of multimedia forensics. The results of image source identification can serve as a minimum reference in the field of image forensics, used for screening and assisting in the identification of suspicious images, and can also help trace the source of images. It is also of great significance in holding individuals accountable for illegal images and ensuring the security and reliability of data information.

[0070] ISI technology includes model recognition and individual identification of image acquisition devices. Through model recognition, the identified image acquisition device information can indicate the type of image acquisition device, such as its model number. Figure 1 As shown, by identifying the individual, the determined image acquisition device information can indicate the device that acquired the image. This individual identification of the image acquisition device can also be called source camera identification (SCI).

[0071] The design and manufacturing process of lenses inevitably introduce distortion and aberrations. Distortion and aberrations are both traces left by the lens in the image. Digital cameras have two common types of distortion: chromatic aberration (CA) and spherical aberration (SA). Different wavelengths of light have different refractive indices when passing through a lens; this phenomenon is called chromatic aberration. Light beams emitted from an on-axis object point no longer converge at a single point after refraction by a spherical surface; this phenomenon is called spherical aberration.

[0072] Sensor pattern noise is another important acquisition artifact during the image acquisition phase. Among various sensor noises, noise caused by photo response nonuniformity (PRNU) is particularly significant.

[0073] PRNU (Photosensitive Numerical Unit) is a characteristic closely related to the imaging sensor in image acquisition equipment. The formation of PRNU is due to variations in the thickness of the silicon on the photosensitive element during sensor manufacturing. Photosensitive elements can be, for example, complementary metal-oxide-semiconductor (CMOS) or charge-coupled devices (CCD). This difference in silicon thickness is a subtle, random defect, and different image acquisition devices will leave different PRNUs in the images they capture. Analyzing the image PRNU can provide clues about the image acquisition device.

[0074] Each device has its specific PRNU mode, which can be estimated by processing a large number of images captured by the device. In fact, due to unavoidable errors in sensor manufacturing and the non-uniformity of sensor units, even under the same light intensity, the pixel brightness will vary slightly.

[0075] The image I acquired by the image acquisition device can be represented as: I = (1 + K)I 0 +θ, where K is the PRNU mode of the image acquisition device, I 0Let θ be the original image, and θ be other noise. The PRNU mode is unique for each image acquisition device, is time-stable, and exists in all images acquired by the image acquisition device.

[0076] N images I acquired by image acquisition device i can be used as a basis. i,1 ,...,I i,N The PRNU mode of the image acquisition device is calculated, where N is a positive integer greater than 1. Image I in these N images... i,n noise residual W i,n It can be represented as:

[0077] W i,n =I i,n -f(I i,n )

[0078] Where f() is the denoising algorithm.

[0079] PRNU mode K of image acquisition device i i It can be represented as W i,1 ,...,W i,N Average value:

[0080]

[0081] Alternatively, the N images I acquired by image acquisition device i can be denoised and subjected to maximum likelihood estimation, etc. i,1 ,...,I i,N Processing is performed to obtain the PRNU mode K of the image acquisition device i. i .

[0082] To compare a certain image I m Whether it belongs to this PRNU mode, that is, whether it was acquired by this image acquisition device i, can be used to calculate image I. m noise residual W m PRNU mode K with image acquisition device i i The correlation between them (normalization cross-correlation, NCC) NCC(W) m ,K i ):

[0083]

[0084] Where < > denotes the inner product, and || || denotes the Euclidean norm. Calculate image I. m noise residual W m Image I can be represented m PRNU mode.

[0085] The peak-to-correlation energy ratio (PCE) can be used to determine whether an image under test originates from a specific digital image sensor. After calculating the PCE value, it is compared with a threshold. If the PCE value is greater than or equal to the threshold, the image under test is considered to originate from a digital image sensor with that PRNU noise.

[0086] By calculating image I m noise residual W m PRNU mode K of image acquisition device i i By analyzing the correlation between images, we can determine the relationship between them. m Whether the image was captured by image acquisition device i.

[0087] Alternatively, in determining the computational image I m noise residual W m PRNU mode K with image acquisition device i i Then, a two-channel convolutional neural network was used to process image I. m noise residual W m PRNU mode K with image acquisition device i i The process is performed to obtain the recognition score c. m Fraction c m This is directly related to the consistency between the image and the device. In other words, it depends on the recognition score c. m Image I can be determined m Whether it was acquired by image acquisition device i.

[0088] Denoising algorithms are effective for images with small color variations, i.e., small color gradients. Therefore, N solid-color images (e.g., images of solid-color walls, skies, etc.) can be selected from the images acquired by image acquisition device i as image I. i,1 ,...,I i,N .

[0089] Using the above method, when determining the acquisition device of a certain image among multiple acquisition devices, it is necessary to compare the acquisition trace information of the image with the device acquisition trace information of the multiple acquisition devices, which is a relatively complex process.

[0090] Furthermore, the PRNU mode of the acquisition device needs to be determined during the processing. The accuracy of the determined PRNU mode is significantly affected by the post-processing methods used on the selected image, as well as the complexity of the image content and color extraction. This, in turn, impacts the accuracy of the image source identification results.

[0091] To address the aforementioned problems, this application provides an image processing method and apparatus.

[0092] Figure 2 This is a schematic flowchart illustrating a data processing method provided in an embodiment of this application. The data processing method provided in this embodiment can be specifically applied to an electronic device that implements data processing functions. This electronic device can be a terminal device, a server, etc. In other embodiments, the electronic device can also be other devices; however, this embodiment does not specifically limit the application to these other methods.

[0093] like Figure 2 The data processing methods shown include S201 to S202.

[0094] In S201, first acquisition trace information of the image to be processed is obtained. The first acquisition trace information is used to indicate the acquisition trace formed in the image to be processed by the target acquisition device when acquiring the image to be processed.

[0095] Due to manufacturing variations in image acquisition equipment, acquisition traces will be left in the acquired images. The first acquisition trace information is used to indicate the acquisition traces left by the target acquisition device in the image to be processed.

[0096] In S202, the first acquisition trace information is classified using a classification module to determine the target acquisition device information for acquiring the image to be processed. The classification module is a trained neural network model.

[0097] pass Figure 2 The method shown eliminates the need to sequentially compare the image acquisition trace information with the device acquisition trace information of the multiple acquisition devices when determining the acquisition device of an image from multiple acquisition devices. This reduces processing complexity and improves processing efficiency.

[0098] In S201, a denoising algorithm can be used to determine the first acquisition trace information of the image to be processed.

[0099] Alternatively, the trace extraction module can be used to extract the first acquisition trace information of the image to be processed.

[0100] The trace extraction module may include a trace extraction network for extracting first acquisition trace information from the image to be processed. Alternatively, the trace extraction module may call the trace extraction network through an interface to extract the first acquisition trace information. The trace extraction network is a trained neural network model.

[0101] The trace extraction module and classification module belong to the image processing system. The structure of the image processing system can be found in [reference needed]. Figure 3 and Figure 5 Explanation.

[0102] In some embodiments, the trace extraction module and the classification module can be trained separately.

[0103] A trace extraction module can be trained by collecting trace information using multiple training images and the corresponding labels for each training image. A classification module can be trained by collecting trace information using the labels corresponding to the training images and the information about the label collection device.

[0104] The tag acquisition trace information corresponding to the training image is generally obtained by processing the training image using a denoising algorithm. For post-processed images, or images with large color gradients, the acquisition trace information obtained by the denoising algorithm has low accuracy.

[0105] In other embodiments, the image processing system may be obtained through end-to-end training.

[0106] End-to-end training, also known as end-to-end learning, refers to optimizing the overall goal of a task directly without modular training during the learning process.

[0107] During the end-to-end training of the image processing system, the parameters of the initial image processing system can be adjusted using training data. The training data may include training images and information about the marker acquisition devices corresponding to the training images.

[0108] Compared to training the trace extraction and classification modules separately, training the image processing system end-to-end does not rely on the marker-based trace information corresponding to the training images. Therefore, the image processing system achieves higher accuracy when processing post-processed images or images with large color gradients. In other words, the image processing system trained end-to-end has wider applicability.

[0109] The initial image processing system includes an initial trace extraction module and an initial classification module.

[0110] Before training the initial image processing system, a pre-training process is performed to obtain the initial trace extraction module.

[0111] The initial trace extraction module can be obtained by pre-training using pre-training data, which includes multiple pre-training images and the marker collection trace information corresponding to each pre-training image.

[0112] The marker acquisition trace information corresponding to the pre-training image indicates the acquisition traces formed by the acquisition device that acquired the pre-training image in the pre-training image.

[0113] The initial trace extraction module, which is pre-trained, is used in the initial image processing system, i.e., the pre-trained neural network structure and weights. Therefore, during the training of the initial image processing system, the weights of the initial trace extraction module do not need extensive modification; only fine-tuning is required, thus shortening the training time.

[0114] The marker acquisition trace information corresponding to the pre-trained images can be obtained by processing the pre-trained images using a denoising algorithm. For images with small color changes (i.e., small color gradients) and without post-processing, the acquisition trace information obtained using the denoising algorithm is relatively accurate. In other words, processing post-processed images or images with large color gradients using the initial trace extraction module obtained through pre-training will result in lower accuracy of the acquisition trace information.

[0115] By using training data, the parameters of each module in the initial image processing system are adjusted. The training data can include post-processed images or images with large color gradients. This makes the image processing system widely applicable.

[0116] For details on the training methods of image processing systems, please refer to [link / reference]. Figure 8 and Figure 9 Explanation.

[0117] The image to be processed may include at least one image patch. The size of the at least one image patch may be the same or different. The first acquisition trace information may include the second acquisition trace information of each of the plurality of image patches.

[0118] The more image blocks corresponding to the second acquisition trace information in the first acquisition trace information, and the more diverse the size types of the image blocks, the more accurate the target acquisition device information determined by the classification module.

[0119] These multiple image blocks may or may not cover the entire area of ​​the image to be processed. These multiple image blocks may or may not overlap.

[0120] In the case where the first acquisition trace information includes multiple second acquisition trace information, in S201, the trace extraction module can be used to extract the second acquisition trace information of each image block separately.

[0121] The trace extraction module has the ability to process images of different sizes.

[0122] The extraction of the second acquisition trace information from these multiple image blocks can be performed in parallel, thereby improving processing efficiency.

[0123] For example, the trace extraction module may include multiple trace extraction networks, or the trace extraction module may invoke multiple trace extraction networks, which can perform extraction processing on multiple image patches in parallel to determine the second acquisition trace information of each image patch. The parameters of the multiple trace extraction networks may be the same. The trace extraction network may be a trained neural network model used to extract the acquisition trace information of the image.

[0124] In the case where the first collected trace information includes multiple second collected trace information, in S202, the classification module can be used to perform the following processing:

[0125] Feature extraction processing is performed on the second acquisition trace information of each image block to obtain the feature information of the second acquisition trace information of each image block;

[0126] All the obtained feature information is fused to obtain fused features;

[0127] The fused features are classified to obtain information about the target acquisition device.

[0128] The feature information of the second acquisition trace information of multiple image blocks is fused, and the fused information is used for classification processing to make the identified target acquisition device information more accurate.

[0129] The classification module can include multiple layers and fusion sub-modules.

[0130] The first layer of the plurality of layers is used to extract features from the second acquisition trace information to obtain the feature extraction result of the first layer; the other layers of the plurality of layers are used to extract features from the feature extraction result determined by the previous layer to obtain the feature extraction result of the other layer.

[0131] In other words, each of these multiple layers is used to extract features from the input information to obtain the feature extraction result of that layer. The input information of the first layer is the second acquisition trace information, and the input of the other layers is the feature extraction result of the previous layer.

[0132] The fusion submodule is used to fuse the feature extraction results of multiple layers to obtain the feature information of the second acquisition trace information.

[0133] Neural networks can consist of multiple layers. The initial layers extract low-level semantic features. As the depth of the neural network increases, later layers extract more complex features, such as high-level semantic features. Fusing the feature extraction results from multiple layers allows the feature information to better focus on the local and global features of the second acquisition trace information. By classifying the feature information, the target acquisition device information determined is more accurate.

[0134] The difficulty of brand identification, model identification, and device identification is constantly increasing. In this embodiment of the application, the target acquisition device information can be used to indicate the brand or model of the target acquisition device for acquiring the image to be processed, or it can be used to indicate the target acquisition device for acquiring the image to be processed.

[0135] Figure 3 This is a schematic structural diagram of an image processing system provided in an embodiment of this application.

[0136] The image processing system 300 includes a trace extraction module 310 and a classification module 320.

[0137] The trace extraction module 310 is used to extract the first acquisition trace information of the image to be processed.

[0138] The classification module 320 is used to classify the first acquisition trace information to determine the target acquisition device information of the image to be processed.

[0139] Target acquisition device information can be used to indicate the target acquisition device for acquiring the image to be processed, or it can be used to indicate the type of target acquisition device, such as model, brand, etc.

[0140] In some embodiments, the trace extraction module 310 may use a denoising algorithm to determine the first acquisition trace information of the image to be processed.

[0141] In other embodiments, the trace extraction module 310 may include a trace extraction network for extracting first acquisition trace information of the image to be processed. Alternatively, the trace extraction module 310 may call the trace extraction network through an interface to extract and process the image to be processed in order to obtain the first acquisition trace information of the image to be processed.

[0142] The trace extraction network can be a trained neural network model.

[0143] Neural networks (NNs) can be composed of neural units, also known as processing units or neurons. A neural network is a network formed by connecting multiple such individual neural units together; that is, the output of one neural unit can be the input of another. The input of each neural unit can be connected to the local receptive field of the previous layer to extract features from the local receptive field, which can be a region composed of several neural units.

[0144] A neural unit can refer to a unit of x s The arithmetic unit that takes an intercept of 1 as input can output the following:

[0145]

[0146] Where s = 1, 2, ..., n, n is a natural number greater than 1, W s For x s The weights are denoted by b, where b is the bias of the neural unit. f is the activation function of the neural unit, used to introduce nonlinear characteristics into the neural network to convert the input signal in the neural unit into the output signal. The output signal of this activation function can be used as the input of the next convolutional layer, and the activation function can be the sigmoid function.

[0147] The trace extraction module 310 can use a trace extraction network to process all or part of the region in the image to be processed in order to obtain the first acquisition trace information.

[0148] The first acquisition trace information may include sensor mode noise of the target acquisition device, such as PRNU. The first acquisition trace information may also include distortion and aberrations inevitably introduced by the design and manufacturing process of the lens of the target acquisition device used to acquire the image to be processed.

[0149] The classification module 320 can use a feature extraction network to extract features from the first acquisition trace information to obtain feature information. The classification module 320 can also use a classifier to classify the feature information to obtain target acquisition device information. The structure of the feature extraction network can be found in [reference needed]. Figure 4 Explanation.

[0150] Figure 4 This is a schematic structural diagram of a feature extraction network provided in an embodiment of this application.

[0151] The feature extraction network 400 can be a convolutional neural network (CNN). For example, the feature extraction network 400 can be a deep neural network (DNN).

[0152] Deep neural networks, also known as multilayer neural networks, can be understood as neural networks with multiple hidden layers. Based on the position of different layers, the internal neural network of a DNN can be divided into three categories: input layer, hidden layer, and output layer. Generally, the first layer is the input layer, the last layer is the output layer, and the layers in between are hidden layers. Layers can be fully connected, meaning that any neuron in the i-th layer is connected to any neuron in the (i+1)-th layer.

[0153] A convolutional neural network can include multiple convolutional layers. For example, such as... Figure 4As shown, the feature extraction network 400 may include convolutional layers 411 to 417.

[0154] A convolutional layer is a layer of neurons in a convolutional neural network (CNN) that performs convolutional processing on the input signal. In a convolutional layer, a neuron may only be connected to a subset of neurons in neighboring layers. A convolutional layer typically contains several feature planes, each composed of rectangularly arranged neural units. Neural units within the same feature plane share weights, which are called the convolutional kernel. Shared weights can be understood as the way image information is extracted regardless of location. The convolutional kernel can be initialized as a matrix of random size, and during the training process of the CNN, it learns appropriate weights. Furthermore, the direct benefit of shared weights is reducing the connections between layers in the CNN, thus lowering the risk of overfitting.

[0155] A convolutional layer can contain multiple convolution operators, also known as kernels. In image processing, these operators act as filters, extracting specific information from the input image matrix. Essentially, a convolution operator is a weight matrix, typically predefined. During convolution, the weight matrix processes the input image pixel by pixel (or two pixels by two pixels, depending on the stride) along the horizontal direction, extracting specific features. The size of this weight matrix should be related to the image size. Importantly, the depth dimension of the weight matrix is ​​the same as the depth dimension of the input image; during convolution, the weight matrix extends to the entire depth of the input image. Therefore, convolution with a single weight matrix produces a single-depth convolutional output. However, in most cases, multiple weight matrices of the same size (rows × columns) are used instead. The outputs of each weight matrix are stacked to form the depth dimension of the convolutional image; this dimension can be understood as being determined by the "multiple" dimensions mentioned above. Different weight matrices can be used to extract different features from an image. For example, one weight matrix can be used to extract edge information, another to extract specific colors, and yet another to blur unwanted noise. These multiple weight matrices have the same size (rows × columns), and the resulting convolutional feature maps are also of the same size. These extracted convolutional feature maps are then merged to form the output of the convolution operation.

[0156] The weight values ​​in these weight matrices need to be obtained through extensive training in practical applications. The weight matrices formed by the weight values ​​obtained through training can be used to extract information from the input image, thereby enabling the convolutional neural network to make correct predictions.

[0157] When a convolutional neural network has multiple convolutional layers, the initial convolutional layers (e.g., 411) tend to extract more general features, which can also be called low-level features. As the depth of the convolutional neural network increases, the features extracted by later convolutional layers (e.g., 417) become more and more complex, such as high-level semantic features. Features with higher semantic levels are more suitable for the problem to be solved.

[0158] The feature information can be the features output by the last convolutional layer 417 among the multiple convolutional layers 411 to 417 of the feature extraction network 400.

[0159] Alternatively, the feature extraction network 400 may also include a fusion submodule 420. The fusion submodule 420 is used to fuse features from the outputs of multiple layers in convolutional layers 411 to 417.

[0160] In other words, feature information can be obtained by fusing the features output by multiple convolutional layers in a feature extraction network. For example... Figure 4 As shown, the first acquisition trace information of the image to be processed is input into the feature extraction network 400. The convolutional layers 411 to 417 in the feature extraction network 400 perform convolution processing to obtain the feature information of the first acquisition trace information. The feature information can be obtained by fusing the features output by convolutional layers 411, 414, and 417.

[0161] Convolutional layers 411 to 414 can perform convolution operations with a stride of 2 to achieve downsampling. Downsampling can be understood as reducing the size of the image.

[0162] In convolutional layers 411 to 414, the number of channels in the feature map can increase in multiples of 2. The feature map is one output form of the convolutional layer. The number of channels in a feature map can also be understood as the number of channels in the layer that outputs that feature map. The number of channels in each layer is preset. For example, a black and white image is single-channel and one-dimensional; a color image has three channels (red, green, and blue, i.e., RGB) and is three-dimensional.

[0163] For example, the kernel size of convolutional layers 411 to 414 is 3; convolutional layer 411 has 1 input channel and 64 output channels; convolutional layer 412 has 64 input channels and 128 output channels; convolutional layer 413 has 1 input channel and 64 output channels; convolutional layer 414 has 128 input channels and 256 output channels; and convolutional layers 415 to 417 each have 512 input and output channels. Convolutional layers 415 to 417 can perform convolution operations with a stride of 1.

[0164] The outputs of convolutional layers 411, 414, and 417 can be concatenated after undergoing average pooling operations. In other words, the average pooling results of the outputs of convolutional layers 411, 414, and 417 can be combined.

[0165] For example, the outputs of convolutional layers 411, 414, and 417 can be concatenated after being subjected to global average pooling (GAP) operations, respectively, to achieve the fusion of the output features of convolutional layers 411, 414, and 417.

[0166] Each convolutional layer in the feature extraction network 400 can also include a training acceleration algorithm and an activation function. The training acceleration algorithm could be, for example, BatchNorm. The BatchNorm algorithm can be used to accelerate neural network training, speed up convergence, and improve stability. The activation function could be, for example, the non-linear activation function LeakyReLU.

[0167] Figure 5 This is a schematic structural diagram of another image processing system provided in the embodiments of this application.

[0168] The image processing system 500 includes a trace extraction module 510 and a classification module 520.

[0169] The trace extraction module 510 is used to extract the first acquisition trace information of the image to be processed.

[0170] The trace extraction module 510 includes multiple trace extraction networks. For example... Figure 5 As shown, the trace extraction module 510 includes trace extraction networks 511 to 513.

[0171] Trace extraction networks 511 to 513 can be used to extract traces from different image patches in the image to be processed, in order to determine the second acquisition trace information for each image patch. The image patches processed by trace extraction networks 511 to 513 can be respectively... Figure 5Image blocks 1 to 3 are shown.

[0172] In other words, the first acquisition trace information of the image to be processed includes the second acquisition trace information of each of the multiple image blocks.

[0173] The classification module 520 is used to process the second acquisition trace information of each of the plurality of image blocks to determine the target acquisition device information of the image to be processed.

[0174] The classification model 520 can fuse the multiple second acquisition trace information, extract features from the fused information, and perform classification processing based on the feature extraction results to obtain the target acquisition device information.

[0175] Or, such as Figure 5 As shown, classification model 520 can extract features from each second acquisition trace information to determine the feature information of each second acquisition trace information. Classification model 520 can fuse the feature information of multiple second acquisition trace information to obtain fused information. Classification model 520 can perform classification processing on the fused information to determine the target acquisition device information.

[0176] The classification model 520 may include at least one feature extraction network. Each feature extraction network may extract features from one or more second acquisition trace information.

[0177] Multiple feature extraction networks 521 to 523 can perform feature extraction on the second acquisition trace information in parallel, improving processing efficiency.

[0178] Feature extraction networks 521 to 523 can all be CNNs. Feature extraction networks 521 to 523 can be the same neural network model. For example, the structure of feature extraction network 521 can be the same as the structure of feature extraction network 400. That is, the feature information of the second acquisition trace information can be obtained by fusing features extracted from multiple convolutional layers.

[0179] The classification model 520 may further include a fusion network 524. The fusion network 524 is used to fuse feature information from multiple second-acquisition trace pieces of information to obtain fused information. For example, fusing feature information from multiple second-acquisition trace pieces of information may involve concatenating the feature information from the multiple second-acquisition trace pieces of information.

[0180] The classification model 520 may include a classifier 525. The classifier 525 is used to classify the fused information and determine the target acquisition device information.

[0181] The classifier 525 may include fully connected layers. Each node in a fully connected layer is connected to all nodes in the layer above it. At least one fully connected layer is often placed at the end of a convolutional neural network; this layer transforms the two-dimensional feature map output by the convolutional layers into a one-dimensional vector. The classifier 525 can use the Softmax activation function to classify the output of the fully connected layers.

[0182] The sizes of multiple image patches input into different trace extraction networks can be the same or different.

[0183] During the processing using the image processing system 500, the more image blocks used and the larger the image block size, the more refined the second acquisition trace information, and the more accurate the target device acquisition information obtained by classifying and processing based on the second acquisition trace information.

[0184] Multiple trace extraction networks can be used to process different image patches in parallel to improve processing efficiency.

[0185] Trace extraction networks 511 to 513 can use the same network structure. Trace extraction network 511 uses... Figure 6 The trace extraction network shown Figure 7 The structure of the trace extraction network or other autoencoder network structure shown.

[0186] Figure 6 This is a schematic structural diagram of a trace extraction network provided in an embodiment of this application.

[0187] Figure 6 The trace extraction network shown can employ an encoder-decoder structure, which is a type of convolutional neural network. Figure 6 The trace extraction network shown includes encoder 601 and decoder 602.

[0188] Figure 6 The trace extraction network shown can be a U-Net or other neural network model with an encoder-decoder structure.

[0189] U-Net consists of an encoder and a decoder. The encoder downsamples the input image. The decoder upsamples the result of the encoder's processing. The decoder's output can have the same resolution as the encoder's input image.

[0190] For example, the U-Net encoder can perform four downsampling operations, and the U-Net decoder can perform four upsampling operations.

[0191] Downsampling can be achieved using max pooling layers. Two convolutional layers can be placed between two adjacent max pooling layers.

[0192] Pooling layers reduce the number of training parameters and can be introduced periodically in convolutional neural networks. A pooling layer can follow a convolutional layer, or multiple convolutional layers can be followed by one or more pooling layers. In image processing, the sole purpose of pooling layers is to reduce the spatial size of the image. Pooling layers can include average pooling and / or max pooling operators to sample the input image to obtain a smaller image size. Average pooling calculates the average value of pixel values ​​within a specific range as the result of average pooling. Max pooling takes the pixel with the largest value within a specific range as the result of max pooling. Furthermore, just as the size of the weight matrix in a convolutional layer should be related to the image size, the operators in a pooling layer should also be related to the image size. The size of the output image after pooling can be smaller than the size of the input image of the pooling layer. Each pixel in the output image represents the average or maximum value of the corresponding sub-region of the input image of the pooling layer.

[0193] It should be understood that a pixel value in an image can be a red-green-blue (RGB) color value, which can be a long integer representing the color. For example, a pixel value of 256*Red+100*Green+76Blue, where Blue represents the blue component, Green represents the green component, and Red represents the red component. Within each color component, the smaller the value, the lower the brightness; the larger the value, the higher the brightness. For grayscale images, the pixel value can be a grayscale value.

[0194] Upsampling can be achieved using deconvolution (or up-convolution, up-conv) layers. Two convolutional layers can be placed between two adjacent deconvolutional layers. Two convolutional layers can also be placed between the max-pooling layer used for the final downsampling and the deconvolution layer used for the first upsampling.

[0195] or, Figure 6 The trace extraction network shown can also be obtained by modifying U-Net. For example, to minimize information loss during the trace extraction stage, convolutional layers can be used instead of pooling layers in U-Net to form... Figure 6 The encoder 601 in the trace extraction network is shown. In encoder 601, a convolutional layer replacing the pooling layer is used, with a convolutional kernel of 3×3 and a stride of 2.

[0196] The convolutional layer used to replace the pooling layer serves the same purpose as the replaced pooling layer: downsampling. The modified trace extraction network can have the same number of downsampling operations as U-Net, for example, four.

[0197] In the decoder, deconvolutional layers can be used to reconstruct the image. In the deconvolutional layer, the kernel size is 3×3, and the stride is 2.

[0198] The result of modifying U-Net Figure 6 The trace extraction network shown is a typical fully convolutional neural network. Fully convolutional neural networks have no limitation on the input image size.

[0199] In other words, Figure 6 The trace extraction network shown can process images of different sizes.

[0200] Training with images of fixed scale Figure 6 The trace extraction network shown can process the image to be processed, obtaining the first trace acquisition information of the image. That is... Figure 6 The trace extraction network shown can achieve Figure 3 The function of the trace extraction module 310.

[0201] or, Figure 6 The trace extraction network shown can be used as trace extraction networks 511 to 513 to process image patches of different sizes. That is, the parameters of trace extraction networks 511 to 513 are shared, or it can be understood that the parameters of trace extraction networks 511 to 513 are the same.

[0202] Therefore, there is no need to train each trace extraction network separately, which increases the flexibility of trace extraction network applications and reduces the training time of the trace extraction networks required in the image processing system 500.

[0203] For example, the size and number of multiple image patches in the image to be processed can be flexibly determined based on computing resources. These multiple image patches do not need to cover the entire area of ​​the image to be processed and can be randomly selected. These multiple image patches may or may not overlap.

[0204] For example, if sufficient computing resources are available, the largest possible image patch can be selected.

[0205] Figure 7 This is a schematic structural diagram of a trace extraction network provided in an embodiment of this application.

[0206] Figure 7 The trace extraction network shown is a compressed sensing network (CSNet).

[0207] Figure 7 The trace extraction network shown includes a sampling module 701 and a reconstruction module 702, which can be used to process images.

[0208] The sampling module 701 utilizes the sparsity of data to randomly sample and obtain discrete samples of the image. Discrete samples can be understood as compressed images.

[0209] The reconstruction module 702 uses a nonlinear reconstruction algorithm to reconstruct data from discrete samples.

[0210] The reconstruction module 702 can perform initial reconstruction on discrete samples, initially reconstructing the compressed image. The quality of the initially reconstructed image is generally not high, for example, it may have block artifacts. Subsequently, the reconstruction module 702 can also perform depth reconstruction on the initially reconstructed image. Through depth reconstruction, image enhancement can be achieved, optimizing the initially reconstructed image.

[0211] To ensure good reconstruction results, the sampling rate of sampling module 701 in CSNet can be set to 1. Figure 7 The trace extraction network shown can perform uncompressed sampling and reconstruction.

[0212] Figure 8 This is a schematic flowchart of a training method for an image processing system provided in an embodiment of this application. Figure 8 The training method shown can also be understood as the training method for the neural network model required by the image processing system.

[0213] Figure 8 The training methods shown include S801 and S802.

[0214] In S801, the parameters of the first trace extraction network are adjusted using pre-trained data to obtain the initial trace extraction network.

[0215] The pre-training data includes multiple pre-training images and label acquisition trace information for each pre-training image. The label acquisition trace information for the pre-training images can have the same size as the pre-training images.

[0216] The information on the collected traces can be determined using denoising algorithms such as wavelet transform.

[0217] Denoising algorithms are more accurate for images with small color gradients. In other words, images with small color gradients have less interference with the extraction of acquisition trace information by denoising algorithms, resulting in more reliable extracted acquisition trace information.

[0218] The pre-training images can all be images with color gradients less than a preset value. For example, pre-training images can be images used to record the sky or solid-colored walls. This makes the labeling of collected trace information more accurate.

[0219] For example, a flat-field image can be derived from a pre-training dataset that uses flat-field images from a public dataset (such as the VISION dataset) as pre-training data. Flat-field images have relatively simple scenes and small color gradients.

[0220] Post-processing steps such as image transmission, compression, and decompression affect the accuracy of the denoising results. Pre-training images can all be raw images without post-processing.

[0221] The pre-trained images are input into the first trace extraction network to obtain pre-trained acquired trace information. The parameters of the first trace extraction network are adjusted to reduce the difference between the pre-trained acquired trace information and the labeled acquired trace information. The adjusted first trace extraction network is the initial trace extraction network.

[0222] The initial trace extraction network can also be called a pre-trained trace extraction network.

[0223] The difference between pre-trained acquisition trace information and labeled acquisition trace information can be represented by a loss function.

[0224] In training a deep neural network, to ensure the output closely approximates the desired predicted value, we compare the network's prediction with the target value. Based on the difference, we update the weight vector of each layer (usually pre-configuring parameters before the initial update). For example, if the prediction is too high, the weight vector is adjusted to predict a lower value. This adjustment continues until the deep neural network predicts the target value or a value very close to it. Therefore, we need to predefine "how to compare the difference between the predicted and target values," which is the loss function or objective function. These are important equations used to measure the difference between the predicted and target values. Taking the loss function as an example, a higher output value (loss) indicates a greater difference, and training the deep neural network becomes a process of minimizing this loss.

[0225] The parameters of the first trace extraction network can be adjusted using the backpropagation algorithm.

[0226] Neural networks can employ backpropagation (BP) to correct the parameters of the initial neural network model during training, thereby reducing the reconstruction error loss. Specifically, forward propagation of the input signal to the output generates error loss; this error loss information is then propagated back to update the parameters of the initial neural network model, leading to convergence of the error loss. The backpropagation algorithm is an error-loss-driven backpropagation process aimed at obtaining the optimal parameters of the neural network model, such as the weight matrix.

[0227] The pre-training images can be the same or different sizes. For example, a single image can be used as a pre-training image, or multiple image patches from a single image can be used as separate pre-training images. These multiple image patches can be the same or different sizes. The multiple pre-training images in the pre-training data can include graphic patches from multiple images.

[0228] The loss function of the initial trace extraction network can be expressed as:

[0229]

[0230] Where G represents the first trace extraction network, θ E These are the parameters of the first trace extraction network, B K It is the j-th pre-trained image I i,j In the image patch, N is the batch size. It is the j-th pre-trained image I i,j The marked acquisition trace information, the j-th pre-trained image I i,j It was collected by the i-th device.

[0231] A complete training run using all the data in the training dataset can be called an epoch.

[0232] In one epoch, a portion of the training data can be used to perform a backpropagation parameter update on the neural network model. This portion of data can be understood as a "batch of data", and the amount of this portion of data can be called the batch size.

[0233] For example, the pre-training images can be 512 pixels × 512 pixels, the batch size can be 24, and the epochs can be 300. The parameters of the initial image processing system can be adjusted using the adaptive moment estimation (Adam) optimizer. The Adam optimizer dynamically adjusts the learning rate of each parameter using the first and second moment estimates of the gradient. The initial learning rate can be 0.0004, and after pre-training begins, the learning rate is reduced to 0.5 times the original value every 30 epochs.

[0234] Using S801, the parameters of the first trace extraction network are adjusted based on the target information of the marked acquisition traces extracted in a traditional manner, so that the output of the first trace extraction network gradually approaches the target, completing the first stage of training (i.e., pre-training). Thus, the pre-trained initial trace extraction network possesses a basic ability, similar to the denoising algorithm, to extract acquisition trace information from images with color gradients less than a preset value. In other words, for images with color gradients less than the preset value, the initial trace extraction network can accurately extract acquisition trace information that is essentially the same as the acquisition trace information determined by the denoising algorithm.

[0235] In S802, the parameters of each network in the initial image processing system are adjusted using the training data to obtain the image processing system.

[0236] The training data includes multiple training images and information about the labeling and acquisition devices for each training image.

[0237] It should be understood that different labeled acquisition device information is used to represent different categories. That is, the multiple training images belong to multiple categories. When the trained image processing system processes the image to be processed, the target acquisition device information of the image to be processed is determined to be one of the different labeled acquisition device information in the training data.

[0238] The training images are input into the initial image processing system to obtain training acquisition device information. The parameters of the initial image processing system are adjusted to reduce the discrepancy between the training acquisition device information and the labeled acquisition device information. The adjusted initial image processing system can then be used as the image processing system.

[0239] The initial image processing system includes an initial trace extraction network, an initial feature extraction network, and an initial classifier.

[0240] The initial trace extraction network is used to extract training data from the training images to obtain training acquisition trace information.

[0241] The initial feature extraction network is used to extract features from the training acquisition trace information to obtain training feature information of the training images.

[0242] The initial classifier is used to classify the training feature information to obtain the training acquisition device information.

[0243] In some embodiments, a training image can be input into an initial trace extraction network to obtain training acquisition trace information of the training image.

[0244] The initial feature extraction network may include multiple layers. The first layer extracts features from the information input to the initial feature extraction network to obtain the training feature extraction result of that layer. The other layers (excluding the first layer) extract features from the training feature extraction result of the previous layer to obtain the training feature extraction result of that layer.

[0245] The training feature extraction result of the last layer of the initial feature extraction network can be used as the training feature information of the training image. Alternatively, the initial feature extraction network may include an initial fusion submodule, which can fuse the training feature extraction results extracted from all or some of the multiple layers to obtain a fused result, which can also be used as the training feature information of the training image.

[0246] Therefore, after passing through S801 and S802, Figure 3 The trace extraction module 310 in the image processing system shown may include an adjusted initial trace extraction network, or the adjusted initial trace extraction network may be invoked via an interface. The classification module 320 may include an adjusted initial feature extraction network and an adjusted initial classifier, or the classification module 320 may invoke the adjusted initial feature extraction network and the adjusted initial classifier via an interface.

[0247] In other embodiments, multiple image patches of the training image can be input into the initial trace extraction network to obtain training image patch acquisition trace information for each image patch. The training acquisition trace information of the training image can include the training image patch acquisition trace information of the multiple image patches.

[0248] The initial feature extraction network can extract features from the trace information collected in each training image patch separately to obtain the training image patch feature information for each patch. The initial feature extraction network can then fuse the feature information from these multiple training image patches to obtain the training feature information for the entire training image.

[0249] The initial feature extraction network may include multiple layers. The first layer extracts features from the information input to the initial feature extraction network to obtain the training feature extraction result of that layer. The other layers (excluding the first layer) extract features from the training feature extraction result of the previous layer to obtain the training feature extraction result of that layer.

[0250] The initial feature extraction network can include multiple layers. It processes the acquisition trace information of multiple training image patches, and can use the training feature extraction result of the last layer as the training image patch feature information for that image patch. Alternatively, the initial feature extraction network can also include an initial fusion submodule, which can fuse the training feature extraction results of all or part of the multiple layers, and the fused result can be used as the training image patch feature information for that image patch. The structure of the initial feature extraction network can be found in [reference needed]. Figure 4 Explanation.

[0251] The initial feature extraction network may also include an initial fusion network, which can fuse the training image patch feature information of the multiple image patches to obtain the training feature information of the training image.

[0252] For example, the image patch size can be 512 (pixels) × 512 (pixels), the batch size can be 24, and the epochs can be 300. The parameters of the initial image processing system can be adjusted using the adaptive moment estimation (Adam) optimizer. The initial learning rate can be 0.0004, and after training begins, the learning rate is reduced to 0.5 times the original rate every 50 epochs.

[0253] When multiple image patches of the training image are input into the initial trace extraction network, and the initial feature extraction network and initial classifier are used to collect trace information from the training image patches output by the initial trace extraction network for subsequent processing, the loss function of the initial image processing system can be expressed as:

[0254]

[0255] Where G' represents the initial trace extraction network, θ' E Here are the parameters of the initial trace extraction network, L is the cross-entropy loss function, and θ is... D H() is a parameter of L. c c represents the training acquisition device information determined by the initial image processing system, and b represents the label acquisition device information. K Image patches from the training image.

[0256] Using S802, the parameters of each network in the initial image processing system are adjusted in an end-to-end training manner to achieve joint optimization of the parameters of each network and complete the second stage of training.

[0257] The label acquisition device information for training images can be understood as classification labels. Different label acquisition device information corresponds to different categories. The image processing system obtained through S801 and S802 can process the image to be processed to determine the category of the image.

[0258] The information about the acquisition device can be used to indicate the type of acquisition device used to acquire the training images. This type of acquisition device can include, for example, the brand and model of the acquisition device. Therefore, the image processing system obtained through steps S801 and S802 can identify the brand or model of the acquisition device used to acquire the image to be processed.

[0259] Alternatively, the information about the acquisition device can be used to indicate the acquisition device of the training images. Then, the image processing system obtained through S801 and S802 can achieve individual identification of the image acquisition device.

[0260] The tasks of brand identification, model identification, and device identification (i.e., individual identification) are becoming increasingly difficult.

[0261] The training data may include multiple training images, which may include post-processed images. Furthermore, the color gradient of each training image is not restricted.

[0262] The acquisition trace information extracted from images using denoising algorithms is inaccurate. By further adjusting the parameters of the pre-trained initial trace extraction network using S802, the network can accurately extract acquisition trace information from post-processed images and images with complex content (images with large color gradients), thus improving the network's versatility and flexibility. Therefore, through... Figure 8 The trace extraction network obtained by the method shown has a higher accuracy in extracting the acquisition trace information than the denoising algorithm.

[0263] The wide range and flexibility of trace extraction networks make image processing systems highly versatile and flexible.

[0264] Table 1 is an example of the embodiments described in this application. Figure 8 The training method shown illustrates the individual recognition accuracy (%) of the image processing system trained using this method. The training data for this image processing system includes information on the labeled acquisition devices, indicating three devices of model 1 and two devices each of models 2 through 5. The training data consists of 100 original images acquired by each of the aforementioned devices, and 300 images each after post-processing 1 and post-processing 2. Post-processing 1 and post-processing 2 can be transmitted, for example, via different instant messaging software.

[0265] Table 1

[0266] Model 2 Model 1 Model 3 Model 4 Model 5 Original image 83.33 85.0 77.5 67.0 94.5 Post-processing 1 92.33 99.0 77.0 69.0 96.0 Post-processing 2 96.36 74.0 77.5 83.0 88.0

[0267] Through the embodiments of this application Figure 8 The image processing system trained using the training method shown has high accuracy for device-level image source recognition.

[0268] Therefore, from S801 to S802, Figure 5 In the image processing system shown, trace extraction networks 511 to 513 can all be adjusted initial trace extraction networks. Feature extraction networks 521 to 523 can all be adjusted initial feature extraction networks. Fusion network 524 can be an adjusted initial fusion network. Classification module 525 is an adjusted initial classifier.

[0269] The parameter-adjusted initial fusion network can be used to fuse the feature information of multiple second acquisition trace information obtained by using trace extraction networks 511-513 to obtain fused information.

[0270] Figure 9 This is a schematic flowchart illustrating a training method for an image processing system provided in an embodiment of this application. The data processing method provided in this embodiment can be specifically applied to an electronic device that implements data processing functions. This electronic device can be a terminal device, a server, etc. In other embodiments, the electronic device can also be other devices; however, this embodiment does not specifically limit its implementation.

[0271] like Figure 9 The data processing methods shown include S901 to S904.

[0272] In S901, the training acquisition trace information of the training image and the label acquisition device information of the training image are acquired.

[0273] The labeled acquisition device information for training images can be used to indicate information related to the acquisition device used to acquire the training images. For example, the labeled acquisition device information can indicate the acquisition device used to acquire the training images, or it can indicate the model or brand of the acquisition device.

[0274] In S902, the training acquisition trace information is classified using the initial classification module to obtain training acquisition device information.

[0275] In S903, based on the difference between the training acquisition device information and the labeled acquisition device information, the parameters of the initial classification module are adjusted to obtain a classification module. The classification module is used to classify the first acquisition trace information of the image to be processed in order to determine the target acquisition device information for acquiring the image to be processed.

[0276] In S904, the image processing system is obtained according to the classification module. The image processing system is used to output target acquisition device information for acquiring the image to be processed based on the input image to be processed.

[0277] An image processing system may include a classification module, or the image processing system may use an interface to call the classification module to determine the target acquisition device information of the image to be processed.

[0278] pass Figure 9 The method shown utilizes the training acquisition trace information and the labeled acquisition device information of the training images to train a classification module. This module can then determine the target acquisition device information based on the first acquisition trace information of the image to be processed. When determining the acquisition device of an image from multiple acquisition devices, it eliminates the need to sequentially compare the image's acquisition trace information with the device acquisition trace information of each of the multiple acquisition devices, thus reducing processing complexity and improving processing efficiency.

[0279] In some embodiments, the training acquisition trace information of the training images can be determined using a denoising algorithm.

[0280] Therefore, in S903, the parameters of the initial classification module can be adjusted only based on the difference between the training acquisition device information and the label acquisition device information.

[0281] In other embodiments, in S901, the training acquisition trace information can be extracted using the trace extraction module.

[0282] The trace extraction module can be a trained neural network model. Trace information can be collected using multiple training images and the corresponding labels for each training image, and the trace extraction module can be trained accordingly.

[0283] Therefore, in S903, the parameters of the initial classification module can be adjusted only based on the difference between the training acquisition device information and the label acquisition device information.

[0284] In other words, the trace extraction module and the classification module can be trained separately.

[0285] In some other embodiments, in S901, the training acquisition trace information can be extracted using the initial trace extraction module.

[0286] In S903, the parameters of the initial trace extraction module and the initial classification module can be adjusted according to the difference between the training acquisition device information and the marker acquisition device information to obtain the trace extraction module and the classification module.

[0287] A loss function can be used to represent the difference between the training acquisition device information and the labeled acquisition device information. The parameters of the initial trace extraction module and the initial classification module are adjusted to minimize the loss function. The adjusted initial trace extraction module can be used as a trace extraction module, and the adjusted initial classification module can be used as a classification module.

[0288] Therefore, in S904, an image processing system can be obtained based on the trace extraction module and the classification module.

[0289] Compared to training the trace extraction module and the classification module separately, the end-to-end training method for the image processing system does not rely on the acquisition trace information of the training image determined by the denoising algorithm, making the image processing system more widely applicable.

[0290] The initial trace extraction module can be obtained by pre-training using pre-training data. The pre-training data includes multiple pre-training images and the marker acquisition trace information corresponding to each pre-training image.

[0291] Pre-trained images can be images that have not undergone post-processing and have small color gradients (e.g., less than a preset value).

[0292] The tag acquisition trace information corresponding to the pre-trained images is generally obtained using denoising algorithms. For post-processed images, or images with large color gradients, the acquisition trace information obtained by denoising algorithms has low accuracy.

[0293] For images with color gradients less than a preset value, the initial trace extraction network obtained after pre-training can accurately extract the acquisition trace information, which is basically the same as the acquisition trace information determined by the denoising algorithm.

[0294] The number of training images can be multiple. These multiple training images may include post-processed images, as well as images with large color gradients, such as those greater than a preset value.

[0295] Therefore, by further adjusting the parameters of the initial trace extraction network obtained through end-to-end training, the trace extraction network can be made to accurately extract the collection trace information from post-processed images and images with complex content (images with large color gradients), thereby improving the versatility and flexibility of image processing system applications.

[0296] Depending on the amount of computing resources available, the trained image processing system can process at least one image block in the image to be processed during the processing of the image to be processed.

[0297] In order to improve the accuracy of the target acquisition device information determined when the image processing system processes multiple image blocks, in S901, the initial trace extraction module can be used to extract the training image block acquisition trace information of each image block in the multiple image blocks of the training image. The training acquisition trace information includes the training image block acquisition trace information of each image block.

[0298] In S902, the initial classification module can be used to perform the following processing:

[0299] Feature extraction processing is performed on the training image block acquisition trace information of each image block to obtain the training feature information of the training image block acquisition trace information of each image block;

[0300] All the obtained training feature information is fused to obtain training fusion features;

[0301] The training fusion features are classified to obtain the training acquisition device information.

[0302] This can improve the accuracy of the image processing system.

[0303] The initial classification module can include multiple layers and an initial fusion submodule.

[0304] Each layer in the initial classification module can be used to extract features from the input information to obtain the training feature extraction result of that layer. Specifically, the input information for the first layer can be the collection trace information of the training image patch, while the input information for other layers can be the training feature extraction result of the previous layer.

[0305] The initial fusion submodule can be used to fuse the training feature extraction results of multiple layers to obtain the training feature information of the acquisition trace information of the training image patch.

[0306] During training, the training feature extraction results of multiple layers are used to extract features from the acquisition trace information of the training image patch to determine the training feature information of the image patch, thereby determining the training acquisition device information of the training image to which the image patch belongs. Correspondingly, during the feature extraction of the second acquisition trace information of the image patch in the image to be processed, the trained image processing system determines the feature information of the second acquisition trace information based on the feature extraction results of multiple layers, thereby determining the target acquisition device information. By comprehensively considering semantics of different levels of complexity, the processing results of the image processing system are made more accurate.

[0307] This application also provides an image source tracing method, including:

[0308] Obtain the image to be traced sent by the client.

[0309] Extract the first acquisition trace information of the image to be traced, the first acquisition trace information being used to indicate the acquisition traces formed in the image to be traced by the target acquisition device when acquiring the image to be traced;

[0310] The first acquisition trace information is classified using a classification module to determine the target acquisition device information for acquiring the image to be traced.

[0311] The source tracing result is sent to the client, and the source tracing result includes the target acquisition device information.

[0312] The image tracing method provided in this application embodiment can realize image tracing and determine the acquisition device information of the image to be traced.

[0313] This application also provides an image identification method, including:

[0314] Acquire the image to be identified and the information of the acquisition device to be determined;

[0315] Extract the first acquisition trace information of the image to be identified, wherein the first acquisition trace information is used to indicate the acquisition trace formed in the image to be identified by the target acquisition device when acquiring the image to be identified;

[0316] The first acquisition trace information is classified using a classification module to determine the target acquisition device information for acquiring the image to be identified.

[0317] A judgment result is sent to the client, which indicates whether the target acquisition device information is consistent with the pending acquisition device information.

[0318] The image identification method provided in this application embodiment can determine whether the acquisition device information of the image to be identified is consistent with the acquisition device information to be determined.

[0319] For details on the processing of images to be traced and identified in image tracing and image identification methods, please refer to [link to relevant documentation]. Figures 2 to 9 Explanation.

[0320] The above text combined Figures 1 to 9 The image processing method, image processing system, and image processing system training method provided in the embodiments of this application are described below. Figures 10 to 11 This section describes the apparatus embodiments of the present application. It should be understood that the descriptions of the image processing method, image processing system, and training method of the image processing system correspond to the descriptions of the apparatus embodiments; therefore, any parts not described in detail can be referred to the above description.

[0321] Figure 10This is a schematic structural diagram of an electronic device provided in an embodiment of this application.

[0322] Figure 10 The electronic device shown includes an acquisition module 1010 and a processing module 1020.

[0323] In some embodiments, the acquisition module 1010 is used to acquire first acquisition trace information of the image to be processed, the first acquisition trace information being used to indicate the acquisition trace formed in the image to be processed by the target acquisition device when acquiring the image to be processed.

[0324] The processing module 1020 is used to classify the first acquisition trace information using the classification module to determine the target acquisition device information for acquiring the image to be processed. The classification module is a trained neural network model.

[0325] Optionally, the acquisition module 1010 is specifically used to extract the first acquisition trace information of the image to be processed using the trace extraction module, wherein the trace extraction module is a trained neural network model.

[0326] The trace extraction module and the classification module belong to the image processing system, which is trained using training data, including training images and information about the marker acquisition devices corresponding to the training images.

[0327] Optionally, the image processing system is obtained by training an initial image processing system using the training data. The initial image processing system includes an initial trace extraction module, which is obtained by pre-training using pre-training data. The pre-training data includes multiple pre-training images and marker acquisition trace information corresponding to each pre-training image. The trace extraction module is obtained based on the initial trace extraction module.

[0328] Optionally, the first acquisition trace information includes the second acquisition trace information of each of the plurality of image blocks in the image to be processed, wherein the plurality of image blocks have different sizes.

[0329] Optionally, the processing module 1020 is also used to: extract the second acquisition trace information of each image block using the trace extraction module.

[0330] Optionally, the processing module 1020 is further configured to: perform the following processing using the classification module:

[0331] Feature extraction processing is performed on the second acquisition trace information of each image block to obtain the feature information of the second acquisition trace information of each image block;

[0332] All the obtained feature information is fused to obtain fused features;

[0333] The fused features are classified to obtain information about the target acquisition device.

[0334] Optionally, the processing module 1020 is further configured to: the classification module includes multiple layers and a fusion submodule;

[0335] The first layer of the plurality of layers is used to extract features from the second acquisition trace information to obtain the feature extraction result of the first layer;

[0336] The layers other than the first layer are used to extract features from the feature extraction results determined by the previous layer, so as to obtain the feature extraction results of the other layers.

[0337] The fusion submodule is used to fuse the feature extraction results of multiple layers to obtain the feature information of the second acquisition trace information.

[0338] Optionally, the target acquisition device information is used to indicate the target acquisition device for acquiring the image to be processed.

[0339] In other embodiments, the acquisition module 1010 is used to acquire training acquisition trace information of the training image and label acquisition device information of the training image.

[0340] The processing module 1020 is used to classify the training acquisition trace information using the initial classification module to obtain training acquisition device information.

[0341] The processing module 1020 is further configured to adjust the parameters of the initial classification module based on the difference between the training acquisition device information and the labeled acquisition device information to obtain a classification module. The classification module is used to classify the first acquisition trace information of the image to be processed in order to determine the target acquisition device information for acquiring the image to be processed.

[0342] The processing module 1020 is further configured to obtain the image processing system according to the classification module, wherein the image processing system is configured to output target acquisition device information for acquiring the image to be processed based on the input image to be processed.

[0343] Optionally, the acquisition module 1010 is specifically used to extract the training acquisition trace information using the initial trace extraction module.

[0344] The processing module 1020 is further configured to adjust the parameters of the initial trace extraction module and the initial classification module based on the difference between the training acquisition device information and the marker acquisition device information, thereby obtaining the trace extraction module and the classification module.

[0345] The processing module 1020 is further configured to obtain the image processing system based on the trace extraction module and the classification module.

[0346] Optionally, the initial trace extraction module is obtained by pre-training using pre-training data, which includes multiple pre-training images and the marker acquisition trace information corresponding to each pre-training image.

[0347] Optionally, the acquisition module 1010 is specifically used to extract training image block acquisition trace information of each image block in multiple image blocks of the training image using the initial trace extraction module, wherein the training acquisition trace information includes training image block acquisition trace information of each image block.

[0348] Processing module 1020 is used to perform the following processing using the initial classification module:

[0349] Feature extraction processing is performed on the training image block acquisition trace information of each image block to obtain the training feature information of the training image block acquisition trace information of each image block;

[0350] All the obtained training feature information is fused to obtain training fusion features;

[0351] The training fusion features are classified to obtain the training acquisition device information.

[0352] Optionally, the marker acquisition device information is used to indicate the training acquisition device that acquired the training images.

[0353] Figure 11 This is a schematic structural diagram of an electronic device provided in an embodiment of this application.

[0354] Figure 11 The electronic device shown includes a memory 1101, a processor 1102, a communication interface 1103, and a communication bus 1104. The memory 1101, processor 1102, and communication interface 1103 are interconnected via the communication bus 1104.

[0355] The memory 1101 may be a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1101 may store a program. When the program stored in the memory 1101 is executed by the processor 1102, the processor 1102 and the communication interface 1103 are used to execute the various steps of the neural network model compression method of the embodiments of this application.

[0356] The processor 1102 may be a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), graphics processing unit (GPU), or one or more integrated circuits, used to execute related programs to implement the functions required by the units in the neural network model compression apparatus of the present application embodiments, or to execute the neural network model compression method of the method embodiments of the present application.

[0357] The processor 1102 can also be an integrated circuit chip with signal processing capabilities. In implementation, each step of the neural network model compression method of this application can be completed by the integrated logic circuits in the hardware of the processor 1102 or by instructions in software form. The processor 1102 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or can be executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 1101. Processor 1102 reads the information in memory 1101 and, in conjunction with its hardware, performs the functions required by the units included in the neural network model compression apparatus of this application embodiment, or performs the neural network model compression method of this application method embodiment.

[0358] The communication interface 1103 uses a transceiver device, such as, but not limited to, a transceiver, to implement... Figure 11 The illustrated electronic device communicates with other devices or communication networks. For example, it can acquire, through communication interface 1103, one or more of the following: first data, second data, the proportion of positive sample data in the second data, parameters of the first neural network model, and PU classifiers.

[0359] Communication bus 1104 may be included in Figure 11The illustrated electronic device shows a pathway for transmitting information between its various components (e.g., memory 1101, processor 1102, communication interface 1103).

[0360] In some embodiments, when the program instructions are executed in the processor 1102, the processor 1102 is configured to:

[0361] Acquire first acquisition trace information of the image to be processed, wherein the first acquisition trace information is used to indicate the acquisition trace formed in the image to be processed by the target acquisition device when acquiring the image to be processed;

[0362] The first acquisition trace information is classified using a classification module to determine the target acquisition device information for acquiring the image to be processed. The classification module is a trained neural network model.

[0363] Optionally, processor 1102 is also used for:

[0364] The first acquisition trace information of the image to be processed is extracted using a trace extraction module, wherein the trace extraction module is a trained neural network model;

[0365] The trace extraction module and the classification module belong to the image processing system, which is trained using training data, including training images and information about the marker acquisition devices corresponding to the training images.

[0366] Optionally, the image processing system is obtained by training an initial image processing system using the training data. The initial image processing system includes an initial trace extraction module, which is obtained by pre-training using pre-training data. The pre-training data includes multiple pre-training images and marker acquisition trace information corresponding to each pre-training image. The trace extraction module is obtained based on the initial trace extraction module.

[0367] Optionally, the first acquisition trace information includes the second acquisition trace information of each of the plurality of image blocks in the image to be processed, wherein the plurality of image blocks have different sizes.

[0368] Optionally, the processor 1102 is also used to: extract second acquisition trace information for each image block using the trace extraction module.

[0369] Optionally, the processor 1102 is further configured to: perform the following processing using the classification module:

[0370] Feature extraction processing is performed on the second acquisition trace information of each image block to obtain the feature information of the second acquisition trace information of each image block;

[0371] All the obtained feature information is fused to obtain fused features;

[0372] The fused features are classified to obtain information about the target acquisition device.

[0373] Optionally, the processor 1102 is further configured to: the classification module includes multiple layers and a fusion submodule;

[0374] The first layer of the plurality of layers is used to extract features from the second acquisition trace information to obtain the feature extraction result of the first layer;

[0375] The layers other than the first layer are used to extract features from the feature extraction results determined by the previous layer, so as to obtain the feature extraction results of the other layers.

[0376] The fusion submodule is used to fuse the feature extraction results of multiple layers to obtain the feature information of the second acquisition trace information.

[0377] Optionally, the target acquisition device information is used to indicate the target acquisition device for acquiring the image to be processed.

[0378] In other embodiments, when the program instructions are executed in the processor 1102, the processor 1102 is configured to:

[0379] Acquire training acquisition trace information of the training images and label acquisition device information of the training images;

[0380] The training acquisition trace information is classified using an initial classification module to obtain training acquisition device information;

[0381] Based on the difference between the training acquisition device information and the labeled acquisition device information, the parameters of the initial classification module are adjusted to obtain a classification module. The classification module is used to classify the first acquisition trace information of the image to be processed in order to determine the target acquisition device information for acquiring the image to be processed.

[0382] The image processing system is obtained according to the classification module, and the image processing system is used to output target acquisition device information for acquiring the image to be processed based on the input image to be processed.

[0383] Optionally, processor 1102 is also used for:

[0384] The training acquisition trace information is extracted using the initial trace extraction module;

[0385] Based on the difference between the training acquisition device information and the label acquisition device information, adjust the parameters of the initial trace extraction module and the initial classification module to obtain the trace extraction module and the classification module;

[0386] The image processing system is obtained based on the trace extraction module and the classification module.

[0387] Optionally, the initial trace extraction module is obtained by pre-training using pre-training data, which includes multiple pre-training images and the marker acquisition trace information corresponding to each pre-training image.

[0388] Optionally, processor 1102 is also used for:

[0389] The initial trace extraction module is used to extract training image block acquisition trace information for each image block in multiple image blocks of the training image. The training acquisition trace information includes training image block acquisition trace information for each image block.

[0390] The initial classification module is used to perform the following processing:

[0391] Feature extraction processing is performed on the training image block acquisition trace information of each image block to obtain the training feature information of the training image block acquisition trace information of each image block;

[0392] All the obtained training feature information is fused to obtain training fusion features;

[0393] The training fusion features are classified to obtain the training acquisition device information.

[0394] Optionally, the marker acquisition device information is used to indicate the training acquisition device that acquired the training images.

[0395] This application also provides a computer program storage medium, characterized in that the computer program storage medium has program instructions, which, when executed, cause the training method of the above-mentioned data processing method or image processing system to be executed.

[0396] This application also provides a chip system, characterized in that the chip system includes at least one processor, and when program instructions are executed in the at least one processor, the above-described data processing method or image processing system training method is executed.

[0397] In this application, prefixes such as "first" and "second" are used merely to distinguish different descriptive objects and do not limit the position, order, priority, quantity, or content of the described objects. For example, if the described object is "state data," then the ordinal number before "collection trace information" in "first collection trace information" and "second collection trace information" does not restrict the position, order, or priority of the "collection trace information."

[0398] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

[0399] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0400] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0401] 1. Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage media, or any other non-transferable media that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include non-transitory computer-readable media, such as modulated data signals and carrier waves.

[0402] 2. Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0403] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. A data processing method, characterized in that, include: Acquiring first acquisition trace information of an image to be processed includes: extracting the first acquisition trace information of the image to be processed using a trace extraction module, wherein the trace extraction module is a trained neural network model; the first acquisition trace information is used to indicate the acquisition traces formed in the image to be processed by the target acquisition device when acquiring the image to be processed; the first acquisition trace information includes second acquisition trace information of each image block in a plurality of image blocks of the image to be processed; the trace extraction module extracts the second acquisition trace information of each image block respectively; the plurality of image blocks are of different sizes, and the plurality of image blocks do not cover the entire area of ​​the image to be processed; The classification module classifies the first acquisition trace information to determine the target acquisition device information for acquiring the image to be processed. This includes: performing the following processing using the classification module: extracting features from the second acquisition trace information of each image block to obtain feature information for each image block; fusing all obtained feature information to obtain fused features; and classifying the fused features to obtain the target acquisition device information. The classification module is a trained neural network model, including multiple layers and a fusion submodule. The first layer of the multiple layers is used to extract features from the second acquisition trace information to obtain the feature extraction result of the first layer. Other layers besides the first layer are used to extract features from the feature extraction result determined by the previous layer to obtain the feature extraction results of the other layers. The fusion submodule is used to fuse the feature extraction results of multiple layers to obtain the feature information of the second acquisition trace information.

2. The method according to claim 1, characterized in that, The trace extraction module and the classification module belong to the image processing system, which is trained using training data. The training data includes training images and information about the marker acquisition devices corresponding to the training images.

3. The method according to claim 2, characterized in that, The image processing system is obtained by training the initial image processing system using the training data. The initial image processing system includes an initial trace extraction module, which is obtained by pre-training using pre-training data. The pre-training data includes multiple pre-training images and the marker acquisition trace information corresponding to each pre-training image. The trace extraction module is obtained based on the initial trace extraction module.

4. The method according to any one of claims 1-3, characterized in that, The target acquisition device information is used to indicate the target acquisition device for acquiring the image to be processed.

5. A training method for an image processing system, characterized in that, The method includes: Acquire training acquisition trace information of the training images and label acquisition device information of the training images; The training acquisition trace information is classified using an initial classification module to obtain training acquisition device information. The initial classification module is a neural network model. Based on the difference between the training acquisition device information and the labeled acquisition device information, the parameters of the initial classification module are adjusted to obtain a classification module. This classification module is used to classify the first acquisition trace information of the image to be processed, in order to determine the target acquisition device information for acquiring the image. The classification module performs the following processing: feature extraction processing is performed on the second acquisition trace information of each image block to obtain feature information of the second acquisition trace information of each image block; all obtained feature information is fused to obtain fused features; the fused features are then classified to obtain the target acquisition device information. The classification module is a trained neural network model, including multiple layers and a fusion submodule; the first layer of the multiple layers is used to perform feature extraction processing on the second acquisition trace information. Feature extraction is performed to obtain the feature extraction result of the first layer; the other layers among the multiple layers, excluding the first layer, are used to extract features from the feature extraction result determined by the previous layer to obtain the feature extraction results of the other layers; the fusion submodule is used to fuse the feature extraction results of multiple layers to obtain the feature information of the second acquisition trace information; the first acquisition trace information is extracted from the image to be processed using the trace extraction module, which is a trained neural network model; the first acquisition trace information includes the second acquisition trace information of each image block in multiple image blocks of the image to be processed; the trace extraction module extracts the second acquisition trace information of each image block respectively; the multiple image blocks are of different sizes and do not cover the entire area of ​​the image to be processed; The image processing system is obtained according to the classification module, and the image processing system is used to output target acquisition device information for acquiring the image to be processed based on the input image to be processed.

6. The method according to claim 5, characterized in that, The step of obtaining training acquisition trace information of training images includes: extracting the training acquisition trace information using an initial trace extraction module, wherein the initial trace extraction module is a neural network model; The step of adjusting the parameters of the initial classification module based on the difference between the training acquisition device information and the label acquisition device information includes: adjusting the parameters of the initial trace extraction module and the initial classification module based on the difference between the training acquisition device information and the label acquisition device information to obtain the trace extraction module and the classification module; The step of obtaining the image processing system based on the classification module includes: obtaining the image processing system based on the trace extraction module and the classification module.

7. The method according to claim 6, characterized in that, The initial trace extraction module is obtained by pre-training using pre-training data, which includes multiple pre-training images and the marker acquisition trace information corresponding to each pre-training image.

8. An electronic device, characterized in that, include: processor; as well as A memory for storing program instructions, which the image processing apparatus executes via the processor to perform the method as described in any one of claims 1-4 or the method as described in any one of claims 5-7.

9. An image source tracing method, characterized in that, include: Obtain the image to be traced sent by the client. Extracting the first acquisition trace information of the image to be traced includes: extracting the first acquisition trace information of the image to be processed using a trace extraction module, wherein the trace extraction module is a trained neural network model; the first acquisition trace information is used to indicate the acquisition traces formed in the image to be traced by the target acquisition device when acquiring the image to be traced; the first acquisition trace information includes the second acquisition trace information of each image block in a plurality of image blocks of the image to be processed; the trace extraction module extracts the second acquisition trace information of each image block respectively; the plurality of image blocks are of different sizes, and the plurality of image blocks do not cover the entire area of ​​the image to be processed; The classification module classifies the first acquisition trace information to determine the target acquisition device information for the image to be traced. This includes: performing the following processing using the classification module: extracting features from the second acquisition trace information of each image block to obtain feature information for each image block; fusing all obtained feature information to obtain fused features; and classifying the fused features to obtain the target acquisition device information. The classification module is a trained neural network model, including multiple layers and a fusion submodule. The first layer of the multiple layers is used to extract features from the second acquisition trace information to obtain the feature extraction result of the first layer. Other layers besides the first layer are used to extract features from the feature extraction result determined by the previous layer to obtain the feature extraction results of the other layers. The fusion submodule is used to fuse the feature extraction results of multiple layers to obtain the feature information of the second acquisition trace information. The source tracing result is sent to the client, and the source tracing result includes the target acquisition device information.

10. An image identification method, characterized in that, include: Acquire the image to be identified and the information of the acquisition device to be determined; Extracting the first acquisition trace information of the image to be identified includes: extracting the first acquisition trace information of the image to be processed using a trace extraction module, wherein the trace extraction module is a trained neural network model; the first acquisition trace information is used to indicate the acquisition traces formed in the image to be identified by the target acquisition device when acquiring the image to be identified; the first acquisition trace information includes the second acquisition trace information of each image block in a plurality of image blocks of the image to be processed; the trace extraction module extracts the second acquisition trace information of each image block respectively; the plurality of image blocks are of different sizes, and the plurality of image blocks do not cover the entire area of ​​the image to be processed; The classification module classifies the first acquisition trace information to determine the target acquisition device information for the image to be identified. This includes: performing the following processing using the classification module: extracting features from the second acquisition trace information of each image block to obtain feature information for each image block; fusing all obtained feature information to obtain fused features; and classifying the fused features to obtain the target acquisition device information. The classification module is a trained neural network model, including multiple layers and a fusion submodule. The first layer of the multiple layers is used to extract features from the second acquisition trace information to obtain the feature extraction result of the first layer. Other layers besides the first layer are used to extract features from the feature extraction result determined by the previous layer to obtain the feature extraction results of the other layers. The fusion submodule is used to fuse the feature extraction results of multiple layers to obtain the feature information of the second acquisition trace information. A judgment result is sent to the client, which indicates whether the target acquisition device information is consistent with the pending acquisition device information.

Images