Small sample camera source method for collaborative selection of pseudo labels using multi-distance metrics

Abstract

A method for tracing the source of a small sample camera using collaborative pseudo-label selection with multiple distance metrics is disclosed, belonging to the field of image forensics. Technical solution: Feature extraction is performed on both the test sample set and the small sample set; after processing by a multiple distance metric module, a pseudo-label sample set selected based on interpretable features is obtained; the small sample set and the pseudo-label sample set selected based on interpretable features are iteratively updated by a collaborative pseudo-label selection module; the test sample set and the iteratively updated pseudo-label sample set are then processed by a classification model to finally obtain the source camera identification result. Beneficial effects: This invention uses multiple distance metrics to pre-expand the small sample dataset; it uses a collaborative pseudo-label selection method to iteratively update the positive and negative pseudo-label subsets with less noise, wherein the collaborative attention block enables the model to focus on information interactions with long-distance features in the space, thereby effectively solving the problem of insufficient information in the small sample set.

Description

Technical Field

[0001] This invention belongs to the field of image forensics technology, and in particular relates to a camera source tracing method for solving the problem of small sample sizes. Background Technology

[0002] With a plethora of user-friendly image editing applications on the market, authenticating digital images to detect and prevent malicious tampering is crucial. Passive forensics plays a vital role in image forensics because reliable digital watermarks or tags are often difficult to obtain in real-world forensic scenarios. Source camera identification (SCI) is an interesting problem in digital forensics, aiming to identify the source device of a target image. Due to inherent differences between device hardware (e.g., sensors, processors) and software (e.g., built-in image generation algorithms in the image creation process), traces may be left in camera-generated images.

[0003] Studies have shown that existing methods achieve high recognition accuracy when the training set is large enough. However, insufficient training sets (i.e., small sample sets) can significantly impact performance. It is worth noting that constructing large labeled sample sets is expensive and time-consuming (and sometimes even impossible). This problem is closely related to the few-shot problem in the context of machine learning, and typical solutions include data augmentation, data augmentation, and semi-supervised learning. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a small-sample camera source tracing method. It uses multiple distance metrics to pre-expand the small-sample dataset and uses a collaborative pseudo-label selection method to iteratively update the positive and negative pseudo-label subsets with less noise. The collaborative attention block enables the model to focus on information interactions with long-distance features in the space, thereby effectively solving the problem of insufficient information in the small-sample dataset.

[0005] The technical solution is as follows:

[0006] A few-shot camera source tracing method utilizing multi-distance metrics for collaborative selection of pseudo-labels, comprising the following steps:

[0007] S1. Extract features from the test sample set and the small sample set respectively;

[0008] S2. After processing by the multiple distance metric module, a pseudo-label sample set based on interpretability features is obtained;

[0009] S3. The small sample set and the pseudo-label sample set selected based on interpretability features are iteratively updated via the collaborative pseudo-label selection module;

[0010] S4, the test sample set and the iteratively updated pseudo-label sample set are processed by the classification model to finally obtain the source camera recognition result;

[0011] The multi-distance metric module performs the following steps: After extracting the effective features of the samples, multiple distance metrics are used to monitor the selection of pseudo-labels; for each dimension of the sample feature vector, the Euclidean distance, Manhattan distance, and Chebyshev distance between each pair of test and training samples are calculated; by sorting the above distance metrics, m nearest unlabeled samples are selected for each training sample, and the frequency of each sample is counted; then n samples are selected for each class and set as effective pseudo-labeled samples (m, n are positive integers).

[0012] Furthermore, source identification performance is improved by using multiple distance metrics to collaboratively select pseudo-labels. The multiple distance metric module specifically performs the following steps:

[0013] Step 1: Extract multiple feature vectors for all training samples;

[0014] Step 2: Calculate multiple distance parameters: Select i distance parameters, i∈M, and s distance parameters, s∈{S} l :l∈L}}, generate a set of distance vectors Ascending order D s,i And select the first m vector sets, denoted as D. s,i′ Where M represents the set of multiple distance metrics, L represents the set of labels, and S represents the set of labels. l This represents a labeled sample set, and m represents the distance metric threshold.

[0015] Step 3: Select pseudo-labeled samples: Select u unlabeled samples, u∈U, and l labeled samples, l∈L, for the set {D′ s,i :s∈S l For each sample u in {i∈M}, count the number of times sample u is counted; if the number of times sample u is counted is greater than the set count threshold t, then u is labeled with the pseudo-label l of the current category.

[0016] Furthermore, the collaborative pseudo-label selection module specifically performs the following steps:

[0017] The information of the noise labels in the positive sample learning method is fully utilized by learning from negative samples; Equations (1) and (2) are the cross-entropy loss functions of the positive sample learning method and the negative sample learning method, respectively:

[0018]

[0019]

[0020] Where y k Indicates the real label, p kIt is the probability prediction distribution of the model output after passing through the softmax output layer. represents the complementary label in the negative sample learning method, and c represents the number of samples;

[0021] For each sample u∈U in the unlabeled sample set, the corresponding pseudo-label vector g is determined based on the confidence threshold of the positive and negative labels. (i) :

[0022]

[0023] Where τ p and τ n τ represents the confidence threshold for positive and negative labels, respectively. p ,τ n ∈(0,1), This represents the confidence probability of predicting the label; if the probability prediction score is high enough... Then select the positive label; otherwise, the probability prediction score is low enough. Then a negative label is selected; 1 indicates that the probability prediction score meets the confidence threshold, otherwise it is 0;

[0024] Pseudo-label vectors based on high confidence Filter out noise labels with insufficient confidence, thereby effectively reducing noise interference during training;

[0025] The cross-entropy loss based on the confidence threshold of positive and negative labels is modified to Equation (4):

[0026]

[0027] Where s (i) This represents the number of pseudo-labels selected for sample i. This represents the model's probability prediction output. These are pseudo-labels assigned to unlabeled samples; the reliability of pseudo-label samples is further improved by using a subset of pseudo-labels with less noise for iterative training.

[0028] Furthermore, using CFA image statistical features, the following steps are performed:

[0029] For each pixel (m,n) and its neighborhood (2k+1)×(2k+2), the interpolation model is shown in Equation (5):

[0030]

[0031] in, and These represent the weights of the CFA interpolation coefficients for the green, red, and blue channels in a color image, respectively. This represents the α interpolation coefficients near pixel k in the green channel. and Let α represent the interpolation coefficients near pixel k in the red and blue channels, respectively. Based on the color distribution of Bayer CFA, solve for the interpolation coefficients of R and B sampling points, as well as the interpolation coefficients of R and B sampling points at two G sampling points, and calculate their mean and variance to obtain the interpretable features of the CFA interpolation coefficient feature matrix.

[0032] Furthermore, source identification performance is improved by using multiple distance metrics to collaboratively select pseudo-labels. The multiple distance metrics use Manhattan distance, Euclidean distance, and Chebyshev distance to calculate the spatial similarity between CFA features of samples. Pseudo-labels are assigned by finding points with similar features. The expressions of the above three metrics in two-dimensional space are shown in equations (6), (7), and (8).

[0033] c = |x1 - x2| + |y1 - y2| (6)

[0034]

[0035] c = max(|x2-x1|,|y2-y1|) (8)

[0036] Where c represents the measurement result in two-dimensional space, and x1,x2 and y1,y2 represent the horizontal and vertical coordinates of multiple sample features in two-dimensional space.

[0037] Furthermore, in the collaborative pseudo-label selection module, the network's ability to learn features is enhanced through a collaborative attention module: the coordinate attention block gathers features of the input image along two spatial directions to generate a pair of orientation-aware feature maps. The orientation-aware feature maps are weighted using a shared convolutional kernel, and then weighted using different convolutional kernels in the two spatial directions. The attention weight of each feature point is obtained by multiplying the attention weights in the two spatial directions. During the operation of the coordinated attention block, each feature point obtains a global attention weight and attention weights along the two spatial directions to achieve interaction of attention information in the global and local spatial directions.

[0038] The beneficial effects of this invention are:

[0039] The small-sample camera source tracing method described in this invention is a collaborative pseudo-label selection method based on multiple distance metrics, which has the following advantages compared with the prior art:

[0040] 1. To address the issue that deep learning methods based on coordinate pseudo-label selection perform poorly on small sample datasets, this invention uses multiple distance metrics to pre-expand small sample datasets;

[0041] 2. The collaborative pseudo-label selection method is used to iteratively update the positive and negative pseudo-label subsets with less noise. The collaborative attention block enables the model to focus on information interactions with long-distance features in the space, thereby effectively solving the problem of insufficient information in small sample sets. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0043] Figure 1 This is an overall block diagram of the MDM-CPS method of the present invention;

[0044] Figure 2 This is an overall block diagram of feature extraction and multiple distance metrics in this invention;

[0045] Figure 3 This is an overall block diagram of the collaborative pseudo-label selection module of the present invention;

[0046] Figure 4 This is a schematic diagram illustrating the relationship between the accuracy on multiple datasets and the multiple distance metric threshold m and the pseudo-label number threshold n in this invention;

[0047] Figure 5 The results of different methods in this invention are shown in the diagram (Dresden dataset);

[0048] Figure 6 The results of different methods in this invention are shown in the VISION dataset.

[0049] Figure 7 This is a schematic diagram of the stability experiment results of the Dresden dataset of this invention, where the number of samples in each category is 1.

[0050] Figure 8 This is a schematic diagram of the stability experiment results for the VISION dataset of this invention, where the number of samples in each category is 1.

[0051] Figure 9 This is a schematic diagram illustrating the traceability accuracy of the complete camera category in this invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] The following is in conjunction with the appendix Figure 1-9This paper further explains the method for tracing the origins of small-sample cameras by using multiple distance metrics to collaboratively select pseudo-labels.

[0054] To fully utilize information from small samples, this invention proposes a small-sample source camera recognition method based on collaborative pseudo-label selection using multiple distance metrics, termed the MDM-CPS method. The complete framework diagram of the MDM-CPS method is shown below. Figure 1 As shown. This method is based on semi-supervised learning. First, it expands the labeled dataset using multiple distance metrics, and then iteratively updates the labeled dataset after selecting coordinate pseudo-labels. The goal of this invention is to achieve higher source camera recognition accuracy by expanding the data through adding reliable pseudo-label samples to a small sample dataset.

[0055] Phase 1: Multiple Distance Metric Module

[0056] Figure 2 A framework diagram of the feature extraction and multi-distance module is shown to extend the small sample dataset. After extracting the effective features of the samples, multiple distance metrics are used to monitor the selection of pseudo-labels. For each dimension of the sample feature vector, the Euclidean distance, Manhattan distance, and Chebyshev distance between each pair of test and training samples are calculated. By sorting these distance metrics, a number of (e.g., m = 10) nearest unlabeled samples are selected for each training sample, and the frequency of each sample is counted; then, n samples (n ≥ 10) are selected for each class and set as effective pseudo-labeled samples. The pseudocode corresponding to the above process is shown in Algorithm 1.

[0057]

[0058]

[0059] To obtain more prior information when the number of lens samples is limited, CFA image statistical features were used. Details will be described in the algorithm section.

[0060] As mentioned earlier, due to insufficient data in small-sample scenarios, deep learning algorithms cannot be fully trained, which affects the accuracy of source camera recognition. To address this issue, the small-sample dataset is augmented using multiple distance metrics to provide sufficient training samples for subsequent collaborative pseudo-label selection. However, it's important to note that due to the limited number of labeled samples, augmenting with too many pseudo-label samples may lead to issues such as incorrect pseudo-labels and negatively impact model performance. To reduce noise during training and improve overall performance, we limit the number of pseudo-label samples for each class to ensure the accuracy of pseudo-label selection.

[0061] Phase Two: Collaborative Pseudo-Label Selection Module

[0062] After expanding by at least n samples for each camera category, pseudo-label samples were obtained after filtering based on interpretability features and multiple distance metrics. These pseudo-label samples were then added to a small sample set to form a new training sample set. Next, an uncertainty-aware pseudo-label selection method was used for semi-supervised learning. Figure 3 The diagram shows the block diagram of the collaborative pseudo-label selection method, in which the collaborative attention block introduces spatially distant feature information interaction. It uses a subset of positive and negative pseudo-labels with less noise to be updated iteratively, and always selects pseudo-label samples with higher confidence probabilities in each iteration, thereby improving the overall performance of the model.

[0063] For multi-class classification tasks, traditional learning methods utilize the correct labels of samples. However, in the field of semi-supervised learning, since pseudo-labels are sometimes inaccurate, positive sample-based learning methods may provide some incorrect information. As training progresses, the model gradually fits noisy labels, thus reducing performance. Fortunately, negative sample-based learning methods can make full use of the information from these noisy labels. Equations (1) and (2) respectively show the cross-entropy loss functions for positive sample learning and negative sample learning methods.

[0064]

[0065]

[0066] Where y k It's a real tag, p k It is the probability prediction distribution of the model output after passing through the softmax output layer. In negative sample learning, the complementary label represents the number of samples. For multi-class classification tasks, the complementary label indicates the category to which a training sample does not belong. A sample can have multiple complementary labels. The closer the model's prediction is to the complementary label, the greater the loss of the negative sample learning method, allowing the model to utilize the complementary label information.

[0067] For each sample u∈U in the unlabeled sample set, the corresponding pseudo-label vector g can be determined based on the confidence thresholds for positive and negative labels. (i) :

[0068]

[0069] Where τ p and τ n These are the confidence thresholds (τ) for positive and negative labels, respectively. p ,τ n ∈(0,1)), This represents the confidence probability of predicting the label; if the probability prediction score is high enough... This sample is highly likely to belong to the current category, so choose a positive label; otherwise, if the probability prediction score is low enough... This sample is very likely not to belong to the current category, so a negative label is selected. 1 indicates that this item is 1 when the probability prediction score meets the confidence threshold requirement, and 0 otherwise.

[0070] Pseudo-label vectors based on high confidence Noisy labels with insufficient confidence were filtered out, effectively reducing noise interference during training. The cross-entropy loss based on the confidence threshold of positive and negative labels was modified to Equation 4:

[0071]

[0072] Where s (i) It is the number of pseudo-labels selected for sample i. It is the model's probability prediction output. These are pseudo-labels assigned to unlabeled samples. By using a subset of pseudo-labels with less noise for iterative training, the reliability of the pseudo-label samples is further improved, thus enhancing the overall performance of the model.

[0073] Algorithm-related details:

[0074] CFA characteristics:

[0075] Due to cost considerations, the CMOS / CCD image sensor in a camera only provides one color data (red, green, or blue, i.e., RGB channel) for each pixel. The CFA interpolation algorithm can effectively recover the color information of all three channels. Different camera manufacturers typically use different CFA interpolation algorithms. For each pixel (m,n) and its neighborhood (2k+1)×(2k+2), the interpolation model is shown in equation (5).

[0076]

[0077] in, and These are the CFA interpolation coefficient weights for the green, red, and blue channels in a color image, respectively. These are the α interpolation coefficients near pixel k in the green channel. Similarly, and These represent the corresponding coefficients for the red and blue channels, respectively. Based on the color distribution of the Bayer CFA, the G interpolation coefficients at the R and B sampling points, as well as the R and B interpolation coefficients at two G sampling points, are solved, and their mean and variance are calculated to obtain the interpretable features of the CFA interpolation coefficient feature matrix.

[0078] Semi-supervised learning:

[0079] Semi-supervised learning (SSL) primarily addresses the problem of poor algorithm performance with small sample sizes. This method is characterized by not introducing external information, but rather leveraging small sample labels and certain criteria to introduce unlabeled samples, thereby maximizing the use of information from unlabeled samples obtained through various means. It also expands the finite sample set for better model training.

[0080] Multiple distance metrics:

[0081] The spatial similarity between CFA features of samples is calculated using Manhattan distance (i.e., the first norm), Euclidean distance (i.e., the second norm), and Chebyshev distance (i.e., the infinity norm). False labels are assigned by finding points with similar features. The expressions for these three measures in two-dimensional space are shown in Equations 6, 7, and 8:

[0082] c = |x1 - x2| + |y1 - y2| (6)

[0083]

[0084] c = max(|x2-x1|,|y2-y1|) (8)

[0085] Where c represents the measurement result in two-dimensional space, and x1,x2 and y1,y2 represent the horizontal and vertical coordinates of multiple sample features in two-dimensional space.

[0086] Collaborative attention blocks:

[0087] In camera source identification tasks, the model needs to focus on the inherent fingerprint information of the training images. Therefore, this invention utilizes a coordinated attention module to enhance the network's ability to learn expressive features. By embedding the image's location information into the channels, the model can learn global feature information from the training images to achieve better performance.

[0088] To encourage attention blocks to capture remote interactions with precise location information in space, coordinate attention blocks aggregate features of the input image along two spatial directions to generate a pair of orientation-aware feature maps. These orientation-aware feature maps are then weighted using a shared 1×1 convolutional kernel, followed by weighting with different 1×1 convolutional kernels in the two spatial directions. Attention weights for each feature point are obtained by multiplying the attention weights in the two spatial directions. During the coordinated operation of the attention blocks, each feature point receives both global and spatial attention weights, enabling interaction of attention information in both global and local spatial directions.

[0089] Experimental dataset settings:

[0090] To fully validate the effectiveness of the MDM-CPS method, all experiments in this paper used public datasets in the field of image forensics: the Dresden dataset and the VISION dataset, which are some popular image datasets used in research on judicial forensics issues.

[0091] In this experiment, this embodiment selected 14 different categories of devices from the Dresden dataset and 11 different categories of devices from the VISION dataset. Specific device information is shown in Tables 1 and 2. The number of training samples was limited, ranging from 5, 10, 15, 20, and 25 images per class. The test set consisted of 130-438 unlabeled samples per class, totaling 2791 images (in the Dresden dataset) and 2163 images (in the VISION dataset). Based on multiple distance metrics, we selected different numbers of pseudo-labeled samples (10-20) from these unlabeled samples for each class to expand the small sample dataset. Then, a collaborative pseudo-label selection module was used to train the model to obtain the final source camera recognition accuracy results.

[0092] Table 1 Dresden dataset

[0093]

[0094]

[0095] Table 2 VISION Dataset

[0096] Camera model abbreviation Sample size Apple iPad 2 A1 171 Asus_Zenfone2Laser A2 209 Huawei_Ascend H1 155 Lenovo P70A L1 216 LG_D290 L2 227 Microsoft Lumia 640 LTE M1 187 OnePlus A3000 O1 287 Samsung Galaxy S3 S1 207 Sony Xperia Z1 Compact S2 215 Wiko_Ridge4G W1 253 Xiaomi Redmi Note 3 X1 311

[0097] In the multiple distance metric section, to ensure the accuracy of pseudo-labels pre-expanded on a small sample set, it is necessary to match the optimal parameter values ​​for thresholds m and n. This embodiment conducted thorough comparative experiments with 1 and 3 small samples for each class, and with m and n varying from 1 to 50. The results are as follows... Figure 4 As shown, when m = n, the pseudo-label accuracy can approximately reach its maximum point. Furthermore, the smaller the value of n, the fewer pseudo-label samples are ultimately selected, and the higher the accuracy. Considering the balance between the number of pre-expanded pseudo-label samples and the pseudo-label accuracy, we ultimately choose m = n = 10.

[0098] In the comparative experiments, methods using only pseudo-label selection are labeled PS, while those using pseudo-label selection with collaborative attention are labeled CP. Furthermore, to demonstrate the effectiveness of the multiple distance metric method, methods using both multiple distance metric and pseudo-label selection are labeled MDM-PS, and finally, our proposed method is labeled as MDM-CP. The comparative experimental results on multiple datasets are as follows: Figure 5 and Figure 6 As shown.

[0099] Results from multiple experimental datasets demonstrate that each method proposed in this invention shows positive performance improvements, exhibiting similar results across various datasets. Furthermore, the smaller the sample size, the more significant the performance improvement of the small-sample dataset expansion method based on multiple distance metrics, indicating that data expansion is an effective solution for deep learning algorithms in small-sample scenarios. Experimental results demonstrate that our proposed method is broadly applicable to solving the problem of camera recognition with small source samples.

[0100] Meanwhile, to verify the model's stability in environments with very few samples, this invention conducted a series of stability experiments using the proposed MDM-CPS method. When the number of labeled samples in each class is very small (i.e., only one sample per camera class), the quality of the selected samples will have a significant impact on the final classification accuracy. Therefore, 20 sets of randomized experiments were conducted when there was only one small sample per class. The average of the experimental results was taken, as shown below. Figure 7 , 8 As shown. Experimental results demonstrate that the method of this invention can provide reliable source camera recognition performance even with a very small sample size.

[0101] Furthermore, the MDM-CPS method proposed in this invention was compared with other existing methods to verify its performance. The camera levels were the same in all papers to ensure fairness in performance evaluation. Experimental results between this invention and other existing methods are shown in Table 3. The experimental results demonstrate that the experimental results of this invention exhibit better performance than other existing methods.

[0102] Table 3 shows the experimental results compared with other existing methods.

[0103] method Dresden VISION EP 73.84 79.94 MTDEM 75.16 80.49 DEEPSiameseNetWork 85.30 75.20 MDM-CPS 92.43 84.74

[0104] Furthermore, to test the performance of the MDM-CPS method under more complex conditions, a complete set of camera categories was reselected for each dataset, including different camera brands and models, comprising 27 (in the Dresden dataset) and 35 (in the vision dataset) to simulate more complex situations in real-world forensic investigation scenarios. Experimental results are as follows: Figure 9 As shown, this fully demonstrates the powerful performance of the model of the present invention in handling complex multi-classification problems.

[0105] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A small-sample camera source tracing method utilizing multi-distance metric collaborative selection of pseudo-labels, characterized in that, it improves source identification performance by using multi-distance metric collaborative selection of pseudo-labels, and is further characterized in that, The steps are as follows: S1. Extract features from the test sample set and the small sample set respectively; S2. After processing by the multiple distance metric module, a pseudo-label sample set based on interpretability features is obtained; S3. The small sample set and the pseudo-label sample set selected based on interpretability features are iteratively updated via the collaborative pseudo-label selection module; S4, the test sample set and the iteratively updated pseudo-label sample set are processed by the classification model to finally obtain the source camera recognition result; The multi-distance metric module performs the following steps: After extracting the effective features of the samples, multiple distance metrics are used to monitor the selection of pseudo-labels; for each dimension of the sample feature vector, the Euclidean distance, Manhattan distance, and Chebyshev distance between each pair of test and training samples are calculated, and pseudo-labels are selected for each training sample by ranking the above distance metrics. m Find the nearest unlabeled sample and count the frequency of each sample; Then select for each category n 1 sample and set it as a valid pseudo-label sample, where m , n It is a positive integer; The multi-distance metric module specifically performs the following steps: Step 1: Extract multiple feature vectors for all training samples; Step 2: Calculate multiple distance parameters: Select One distance parameter, , One distance parameter, }, generate a set of distance vectors Ascending order and select the first A set of vectors, denoted as ,in Represents a set of multiple distance metrics. Represents a tag set. This indicates a labeled sample set. Indicates the distance metric threshold; Step 3: Select pseudo-label samples: Select One unlabeled sample, , One labeled sample, For sets Each sample in Perform a count; if the sample The number of times the statistics are greater than the set statistical threshold. Then use the pseudo-label of the current category. To mark ; The collaborative pseudo-label selection module specifically performs the following steps: The information of the noise labels in the positive sample learning method is fully utilized by learning from negative samples; Equations (1) and (2) are the cross-entropy loss functions of the positive sample learning method and the negative sample learning method, respectively: (1) (2) in Indicates the true label, It is the probability prediction distribution of the model output after passing through the softmax output layer. This refers to the complementary label in the negative sample learning method. Indicates the number of samples; For each sample in the unlabeled sample set Based on the confidence thresholds of the positive and negative labels, the corresponding pseudo-label vectors are determined. : (3) in and These represent the confidence thresholds for positive and negative labels, respectively. , This represents the confidence probability of predicting the label, if If yes, then select the positive label; otherwise... If the probability prediction score meets the confidence threshold, then a negative label is selected; 1 indicates that the probability prediction score meets the confidence threshold, otherwise it is 0. Pseudo-label vectors based on high confidence This filters out noise labels with insufficient confidence, thereby effectively reducing noise interference during training. The cross-entropy loss based on the confidence threshold of positive and negative labels is modified to Equation (4): (4) in Indicates sample i Number of pseudo-labels selected This represents the model's probability prediction output. These are pseudo-labels assigned to unlabeled samples; the reliability of pseudo-label samples is further improved by using a subset of pseudo-labels with less noise for iterative training.

2. The few-sample camera source tracing method using multi-distance metrics to collaboratively select pseudo-labels as described in claim 1, characterized in that, The CFA image statistical features are used to perform the following steps: For each pixel and its neighborhood The interpolation model is shown in formula (5): (5) in, , and These represent the weights of the CFA interpolation coefficients for the green, red, and blue channels in a color image, respectively. Indicating green channel k Near the pixel Interpolation coefficients, and These represent the red and blue channels respectively. k Near the pixel Interpolation coefficients; based on the color distribution of Bayer CFA, solve for the G interpolation coefficients of R and B sampling points, as well as the R and B interpolation coefficients of two G sampling points, and calculate their mean and variance to obtain the interpretable features of the CFA interpolation coefficient feature matrix.

3. The few-sample camera source tracing method using multi-distance metrics for collaborative selection of pseudo-labels as described in claim 2, characterized in that, The multiple distance metric uses Manhattan distance, Euclidean distance, and Chebyshev distance to calculate the spatial similarity between CFA features of samples; pseudo-labels are assigned by finding points with similar features. The expressions for Manhattan distance, Euclidean distance, and Chebyshev distance in two-dimensional space are shown in equations (6), (7), and (8): (6) (7) (8) in, Represents the measurement result in two-dimensional space. and The horizontal and vertical coordinates represent the features of multiple samples in a two-dimensional space.

4. The few-sample camera source tracing method using multi-distance metric collaborative selection of pseudo-labels as described in claim 2, characterized in that, In the collaborative pseudo-label selection module, the collaborative attention module enhances the network's ability to learn features: the coordinate attention block gathers the features of the input image along two spatial directions to generate a pair of orientation-aware feature maps. The orientation-aware feature maps are weighted with a shared convolutional kernel and then weighted with different convolutional kernels in the two spatial directions. The attention weight of each feature point is obtained by multiplying the attention weights of the two spatial directions. During the coordinated attention block operation, each feature point obtains global attention weights and attention weights along two spatial directions to achieve attention information interaction in global and local spatial directions.

Images