A source camera identification method and device based on multi-scale feature fusion

Abstract

The application discloses a kind of source camera identification method and device based on multi-scale feature fusion, comprising the following steps: based on multi-scale feature fusion, traceability model is constructed, including multi-scale feature extraction module and feature fusion module and classifier;Supervised learning is carried out on traceability model using training image set, and the multi-scale feature extraction module and feature fusion module after extraction supervised learning are extracted into feature extractor;Using feature extractor, the high-dimensional vector of each test image in test image set is extracted, and the high-dimensional vector is used as the fingerprint index of the corresponding source camera class of test image, to build camera fingerprint library;Using feature extractor, the high-dimensional vector of the image to be measured is extracted, based on high-dimensional vector and the fingerprint index in camera fingerprint library, matching retrieval is carried out, and based on the fingerprint index of matching retrieval, the source camera class of the image to be measured is obtained.The method and device can realize efficient source camera identification while ensuring traceability accuracy.

Description

Technical Field

[0001] This invention belongs to the field of source camera recognition technology, specifically relating to a source camera recognition method and apparatus based on multi-scale feature fusion. Background Technology

[0002] In recent years, with the widespread adoption of mobile smart devices, capturing and editing images has become incredibly easy. This has led to the question of how to reliably verify the source and integrity of images. Image forensics is a technique for preventing malicious tampering with images for illegal gain, and source camera identification is one of the hot research topics in the field of digital image forensics, aiming to trace the origin of images.

[0003] Methods for tracing the origin of images mainly revolve around the image generation process: light signals pass through a lens, are converted into electrical signals by a photosensitive element, and then undergo a series of image processing techniques, such as color interpolation, white balance, gamma correction, edge smoothing, and JPEG compression, ultimately resulting in a digital image. Limited by manufacturing processes and algorithmic models, these steps leave inherent characteristics in the digital image, allowing for traceability even if the device brand and model are identical. The method of analyzing sensor mode noise to trace the origin of camera sensors was first proposed by Lukás and Goljan et al. Sensor mode noise is random noise introduced by digital cameras or other imaging devices during image capture due to inherent sensor performance limitations. Theoretical and experimental evidence has shown that, analogous to using fingerprints to track suspects in forensic identification and ballistic detection to trace firearms, the sensor mode noise contained in an image can be used as the unique fingerprint of its imaging device, also known as a camera fingerprint.

[0004] Source camera identification methods based on sensor pattern noise mainly fall into two categories: methods based on traditional digital images and methods based on deep learning. Deep learning-based methods utilize supervised learning and neural network downsampling to transform the image into a high-dimensional vector representation. A softmax layer is then used to obtain the probabilities of different categories, and the category corresponding to the highest probability is selected as the output.

[0005] In 2019, Freire-Obregon D et al. published a paper in Pattern Recognition Letters titled "Deep learning for source camera identification on mobile devices," proposing DavNet. This approach improves the accuracy of source camera identification on mobile devices while reducing computational resource consumption and increasing inference speed. DavNet essentially uses a deep convolutional neural network (CNN) as a feature extractor. The model is trained on a large number of labeled image samples and optimized using Leaky ReLU and Dropout layers, effectively extracting high-level feature representations. Compared to traditional methods, this approach achieves higher source identification accuracy and efficiency. However, existing deep learning-based source camera identification methods only achieve high-precision source identification on small datasets (less than 10 devices). Their performance is mediocre in larger-scale experiments, and with expanded categories, retraining on the complete dataset or fine-tuning on new datasets is required, which consumes significant time and computational resources. Summary of the Invention

[0006] The purpose of this invention is to provide a source camera identification method based on multi-scale feature fusion. By constructing a source tracing model through multi-scale feature fusion, the source camera is identified, and a high-dimensional vector is extracted to construct a camera fingerprint database for matching and retrieval to adapt to the situation of category expansion, thereby achieving efficient device tracing with slight loss of accuracy.

[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a source camera recognition method based on multi-scale feature fusion, comprising the following steps:

[0009] A source tracing model is constructed based on multi-scale feature fusion, including a multi-scale feature extraction module, a feature fusion module, and a classifier.

[0010] The source tracing model is supervised learning using the training image set, and the feature extractor is composed of a multi-scale feature extraction module and a feature fusion module after supervised learning.

[0011] A high-dimensional vector is extracted from each test image in the test image set using a feature extractor, and the high-dimensional vector is used as the fingerprint index of the source camera category corresponding to the test image to construct a camera fingerprint database.

[0012] A high-dimensional vector of the image to be tested is extracted using a feature extractor. The high-dimensional vector is then matched with the fingerprint index in the camera fingerprint database. The source camera category of the image to be tested is obtained based on the matched fingerprint index.

[0013] The multi-scale feature extraction module includes multiple parallel feature extraction branches of different scales. Each feature extraction branch processes the input image through convolution at the corresponding scale, and then feeds it into the residual network after passing through convolutional layers, batch normalization layers, activation layers, and pooling layers to extract deep vectors.

[0014] The residual network includes a pooling layer, a first convolutional block, two second convolutional blocks, an activation layer, and a pooling layer connected in sequence. A first residual connection is constructed between the output of the pooling layer and the input of the first second convolutional block. The first residual connection passes through a convolutional layer and a batch normalization layer. A second residual connection is constructed between the output of the first convolutional block and the input of the second second convolutional block. A third residual connection is constructed between the output of the first second convolutional block and the input of the activation layer and the pooling layer.

[0015] The first convolutional block includes a first-size convolutional layer + batch normalization layer + activation layer, a second-size convolutional layer + batch normalization layer + activation layer, and a first-size convolutional layer + batch normalization layer connected in sequence.

[0016] The second convolutional block consists of an activation layer + pooling layer, a first-size convolutional layer + batch normalization layer + activation layer, a second-size convolutional layer + batch normalization layer + activation layer, and a first-size convolutional layer + batch normalization layer, connected in sequence.

[0017] The feature fusion module concatenates the deep vectors output by each feature extraction branch in the multi-scale feature extraction module to obtain a high-dimensional vector.

[0018] The classifier uses a fully connected layer and a Softmax layer to map and activate the high-dimensional vector output by the feature fusion module to obtain the source camera recognition result.

[0019] The matching retrieval based on high-dimensional vectors and fingerprint indexes in the camera fingerprint database includes: calculating the similarity between the high-dimensional vector and each fingerprint index in the camera fingerprint database, and selecting the fingerprint indexes with high similarity as the fingerprint indexes for matching retrieval.

[0020] Preferably, Euclidean distance, cosine distance, or Chebyshev distance is used to measure the similarity between the high-dimensional vector and each fingerprint index in the camera fingerprint database.

[0021] Secondly, in order to achieve the above-mentioned objectives, the present invention also provides a source camera recognition device based on multi-scale feature fusion, including a model building unit, a supervised learning unit, a camera fingerprint database building unit, and a matching retrieval unit.

[0022] The model building unit is used to build a source tracing model based on multi-scale feature fusion, including a multi-scale feature extraction module, a feature fusion module, and a classifier;

[0023] The supervised learning unit is used to perform supervised learning on the source tracing model using the training image set, and the multi-scale feature extraction module and feature fusion module after supervised learning are used to form a feature extractor.

[0024] The camera fingerprint database construction unit is used to extract the high-dimensional vector of each test image in the test image set using a feature extractor, and to construct the camera fingerprint database using the high-dimensional vector as the fingerprint index of the source camera category corresponding to the test image.

[0025] The matching retrieval unit is used to extract a high-dimensional vector of the image to be tested using a feature extractor, perform matching retrieval based on the high-dimensional vector and the fingerprint index in the camera fingerprint database, and obtain the source camera category of the image to be tested based on the fingerprint index of the matching retrieval.

[0026] Thirdly, to achieve the above-mentioned objectives, embodiments of the present invention also provide a source camera recognition device based on multi-scale feature fusion, including a memory and a processor. The memory is used to store a computer program, and the processor is used to implement the source camera recognition method based on multi-scale feature fusion provided in the first aspect of the present invention when the computer program is executed.

[0027] Fourthly, to achieve the above-mentioned objectives, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, wherein the computer program, when used by a computer, implements the source camera recognition method based on multi-scale feature fusion provided in the first aspect of the present invention embodiments.

[0028] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0029] (1) A source tracing model based on multi-scale feature fusion is used to identify the source camera. Since it can extract deep vectors of multiple scales and splice them into high-dimensional vectors, classification based on high-dimensional vectors can improve the source tracing accuracy of the source camera.

[0030] (2) A camera fingerprint database is constructed using high-dimensional vectors extracted by the feature extractor, and the source camera category is identified by matching retrieval based on the high-dimensional vectors. In this way, when facing the problem of category expansion, the model can be avoided by repeatedly training and fine-tuning while ensuring the accuracy of tracing the source, thus reducing the time cost. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0032] Figure 1 This is a flowchart of the source camera recognition method based on multi-scale feature fusion provided in an embodiment of the present invention;

[0033] Figure 2 This is a flowchart of supervised learning of the source tracing model provided in an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the source tracing model constructed by multi-scale feature fusion provided in an embodiment of the present invention;

[0035] Figure 4 This is a flowchart of the retrieval process based on a multi-scale feature fusion tracing model provided in an embodiment of the present invention;

[0036] Figure 5 This is a schematic diagram of the source camera recognition device based on multi-scale feature fusion provided in an embodiment of the present invention. Detailed Implementation

[0037] 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 of the invention and do not limit the scope of protection of the invention.

[0038] The inventive concept of this invention is as follows: In view of the technical problem that the source camera recognition method based on deep learning in the prior art can only achieve high-precision source tracing of small-scale datasets, and requires retraining when the categories are expanded, which consumes time and computing resources, the embodiments of this invention provide a source camera recognition method and device based on multi-scale feature fusion to improve the source tracing accuracy and efficiency of source camera recognition.

[0039] Figure 1 This is a flowchart of a source camera recognition method based on multi-scale feature fusion provided in an embodiment of the present invention. Figure 1 As shown in the embodiment, a source camera recognition method based on multi-scale feature fusion is provided, including the following steps:

[0040] S1 is a source tracing model built based on multi-scale feature fusion, including a multi-scale feature extraction module, a feature fusion module, and a classifier.

[0041] Multi-scale Fusion Convolutional Neural Network (MFCNN) is an image classification model based on convolutional neural networks. It can extract and fuse features from images at different scales to improve classification accuracy. In this invention, the multi-scale feature extraction module is the focus of the network design. Since the size of the convolutional kernel directly determines the size of the local receptive field, and considering that sensor fingerprints are very weak noise, the designed multi-scale feature extraction module includes multiple parallel feature extraction branches at different scales to capture more complete features. Each feature extraction branch processes the input image through convolution at the corresponding scale, and then feeds it into the residual network after passing through convolutional layers, batch normalization layers, activation layers, and pooling layers to extract deep vectors.

[0042] The residual network includes a pooling layer, a first convolutional block and two second convolutional blocks, and an activation layer + pooling layer connected in sequence. A first residual connection is constructed between the output of the pooling layer and the input of the first second convolutional block. The first residual connection passes through a convolutional layer + batch normalization layer. A second residual connection is constructed between the output of the first convolutional block and the input of the second second convolutional block. A third residual connection is constructed between the output of the first second convolutional block and the input of the activation layer + pooling layer.

[0043] The first convolutional block includes a first-size convolutional layer + batch normalization layer + activation layer, a second-size convolutional layer + batch normalization layer + activation layer, and a first-size convolutional layer + batch normalization layer connected in sequence.

[0044] The second convolutional block includes an activation layer + pooling layer, a first-size convolutional layer + batch normalization layer + activation layer, a second-size convolutional layer + batch normalization layer + activation layer, and a first-size convolutional layer + batch normalization layer, all connected in sequence.

[0045] Specifically, such as Figure 3As shown, an input 64×64 image is processed through three convolutional layers of different sizes: 3×3, 5×5, and 7×7. Then, it passes through a 5×5 convolutional layer, a batch normalization layer, a ReLU activation layer, and a pooling layer before being fed into three parallel residual networks. In each residual network, after the pooling layer, it first enters a first convolutional block consisting of a 1×1 convolutional layer + batch normalization layer + ReLU activation layer, a 3×3 convolutional layer + batch normalization layer + ReLU activation layer, and a 1×1 convolutional layer + batch normalization layer. Then, it passes through two more convolutional blocks consisting of a ReLU activation layer + pooling layer and a 1×1 convolutional layer + batch normalization layer + ReLU activation layer. The second convolutional block consists of a live layer, a 3×3 convolutional layer + batch normalization layer + ReLU activation layer, and a 1×1 convolutional layer + batch normalization layer. Finally, it passes through a ReLU activation layer + pooling layer. A first residual connection is established between the output of the pooling layer and the input of the first second convolutional block, and the first residual connection passes through a 1×1 convolutional layer + batch normalization layer. A second residual connection is established between the output of the first convolutional block and the input of the second second convolutional block. A third residual connection is established between the output of the first second convolutional block and the input of the activation layer + pooling layer. Finally, each residual network outputs a 256-dimensional deep vector.

[0046] The feature fusion module concatenates the deep vectors output from each feature extraction branch in the multi-scale feature extraction module to obtain a high-dimensional vector, such as... Figure 3 As shown, in the feature fusion module, three 256-dimensional deep vectors are concatenated to obtain a 768-dimensional high-dimensional vector.

[0047] The classifier uses fully connected layers and softmax layers to map and activate the high-dimensional vector output by the feature fusion module to obtain the source camera recognition result.

[0048] S2 utilizes the training image set to perform supervised learning on the source tracing model, and extracts the multi-scale feature extraction module and feature fusion module after supervised learning to form a feature extractor.

[0049] Specifically, such as Figure 2 As shown, given a training image set, the input training image is used to identify and output the source camera category through the source tracing model. At the same time, the source camera category label corresponding to the training image is given. The classifier is trained through supervised learning based on cross-entropy loss to optimize the source tracing ability of the source tracing model, so that the source tracing model has a strong sensor pattern noise extraction capability. The multi-scale feature extraction module and feature fusion module after supervised learning constitute the feature extractor, which is used to directly extract the high-dimensional vector of the image.

[0050] S3. Use a feature extractor to extract the high-dimensional vector of each test image in the test image set, and use the high-dimensional vector as the fingerprint index of the source camera category corresponding to the test image to build a camera fingerprint database.

[0051] Specifically, such as Figure 4 As shown, in the generation stage, the test images of each category in the entire test image set are processed through the source tracing model to obtain a 768-dimensional high-dimensional vector, which is then saved as the fingerprint index of the corresponding category, and finally established as a camera fingerprint database.

[0052] S4. Use a feature extractor to extract a high-dimensional vector from the image to be tested. Then, perform a matching retrieval based on the high-dimensional vector and the fingerprint index in the camera fingerprint database. Based on the matching retrieval fingerprint index, obtain the source camera category of the image to be tested.

[0053] Specifically, such as Figure 4 As shown, in the retrieval stage, an image to be tested is input, the source model is used to infer and extract a high-dimensional vector, and the similarity between the high-dimensional vector and each fingerprint index in the camera fingerprint database is calculated using Euclidean distance, cosine distance or Chebyshev distance. The fingerprint index corresponding to the high similarity is selected as the fingerprint index for matching retrieval, and finally the corresponding source camera category is output.

[0054] In summary, a source camera identification method based on multi-scale feature fusion is proposed. This method constructs a source tracing model based on multi-scale feature fusion for source camera identification. Because it can extract deep vectors from multiple scales and concatenate them into high-dimensional vectors, and perform classification based on high-dimensional vectors, it can achieve high-precision source camera identification on large-scale device datasets. At the same time, it uses the high-dimensional vectors extracted by the feature extractor to construct a camera fingerprint database, and performs matching retrieval based on high-dimensional vectors to identify the source camera category. When facing the problem of category expansion, it can avoid repeated training and fine-tuning of the model while ensuring the accuracy of source tracing, thus reducing time overhead.

[0055] Based on the same inventive concept, the embodiments also provide a source camera recognition device 500 based on multi-scale feature fusion, such as... Figure 5 As shown, it includes a model building unit 510, a supervised learning unit 520, a camera fingerprint database building unit 530, and a matching retrieval unit 540;

[0056] Among them, the model building unit 510 is used to build a source tracing model based on multi-scale feature fusion, including a multi-scale feature extraction module, a feature fusion module, and a classifier;

[0057] The supervised learning unit 520 is used to supervise the learning of the source tracing model using the training image set. The feature extractor is composed of the multi-scale feature extraction module and the feature fusion module after supervised learning.

[0058] The camera fingerprint database construction unit 530 is used to extract the high-dimensional vector of each test image in the test image set using the feature extractor, and use the high-dimensional vector as the fingerprint index of the source camera category corresponding to the test image to construct the camera fingerprint database.

[0059] The matching retrieval unit 540 is used to extract a high-dimensional vector of the image to be tested using a feature extractor, perform matching retrieval based on the high-dimensional vector and the fingerprint index in the camera fingerprint database, and obtain the source camera category of the image to be tested based on the fingerprint index of the matching retrieval.

[0060] Based on the same inventive concept, the embodiment also provides a source camera recognition device based on multi-scale feature fusion, including a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to implement the above-described source camera recognition method based on multi-scale feature fusion when the computer program is executed.

[0061] Based on the same inventive concept, the embodiment also provides a computer-readable storage medium storing a computer program, which, when used by a computer, implements the above-described source camera recognition method based on multi-scale feature fusion.

[0062] It should be noted that the source camera recognition device, source camera recognition equipment, and computer-readable storage medium based on multi-scale feature fusion provided in the above embodiments all belong to the same concept as the source camera recognition method embodiments based on multi-scale feature fusion. For details of their specific implementation process, please refer to the source camera recognition method embodiments based on multi-scale feature fusion, which will not be repeated here.

[0063] Experimental Example

[0064] To demonstrate that the present invention outperforms existing methods in terms of source identification accuracy, and to verify the effectiveness of the designed module through self-ablation comparison, an experiment was designed. The dataset was divided into 10 categories of devices (small-scale), 35 categories of devices (medium-scale), and 90 categories of devices (large-scale) according to device size. The data for the 10 categories of devices were selected from the VISION dataset (Shullani D, Fontani M, Iuliani M, et al. Vision: a video and image dataset for source identification[J]. EURASIP Journal on Information Security, 2017, 2017(1):1-16), the data for the 35 categories of devices was the complete VISION dataset, and the data for the 90 categories of devices was the complete Daxing dataset (Tian H, Xiao Y, Cao G, et al. Daxing smartphone identification dataset[J]. IEEE Access, 2019, 7:101046-101053). The models and methods involved include DavNet mentioned above, this invention (3×3), this invention (5×5), this invention (7×7), this invention, and this invention-retrieval. Specifically, this invention (3×3) refers to experiments using only the network branch containing 3×3 convolutional kernels without multi-scale convolutions; this invention (5×5) refers to experiments using only the network branch containing 5×5 convolutional kernels; this invention (7×7) refers to experiments using only the network branch containing 7×7 convolutional kernels; and this invention-retrieval refers to the source tracing model obtained after training the proposed MFCNN on a small-scale dataset of 10 device classes. In the experiments of the three scales, no further training and fine-tuning were performed; instead, the source tracing model was used to directly construct a camera fingerprint database for retrieval to simulate the category expansion scenario. The experimental results are shown in Table 1:

[0065] Table 1. Traceability Testing Results (ACC%)

[0066] Model 10 types of equipment VISION - Class 35 Equipment 90 types of equipment - Daxing DavNet 91.15 78.00 54.00 This invention (3×3) 73.51 91.13 78.94 This invention (5×5) 83.37 90.62 79.43 This invention (7×7) 94.25 91.00 58.11 This invention 98.02 93.20 85.44 This invention - retrieval 97.34 86.90 76.38

[0067] (1) Overall accuracy analysis

[0068] The table shows that the method proposed in this invention achieved the highest source tracing accuracy in all experiments. Thanks to the ability of multi-scale feature fusion to capture comprehensive image features and the feature enhancement capabilities of parallel residual networks, the source tracing model achieved an accuracy of 98.02% in experiments with 10 types of devices; it also achieved 93.2% accuracy on the complete VISION dataset; and even on the complete Daxing dataset, it achieved over 85% accurate source tracing. These data demonstrate that the method proposed in this invention can adapt to source camera identification tasks with a large number of devices.

[0069] In experiments of various scales, this invention-retrieval used the source tracing model trained on 10 types of device data by MFCNN, which achieved the second-best performance of 97.34% accuracy in small-scale experiments; meanwhile, thanks to the good representation of high-dimensional vectors, this invention-retrieval achieved a source tracing accuracy of 76.38% in large-scale experiments, which is close to the best method.

[0070] (2) Category Expansion Analysis

[0071] As the number of devices increases, the source tracing accuracy of almost all models declines significantly, especially DavNet and networks using only 7×7 convolutions, whose accuracy on the Daxing dataset even falls below 60%. For DavNet, this may be due to its relatively simple network structure and insufficient network depth; the original paper only tested this model on a small dataset with 5 devices, without providing in-depth analysis. As for networks using only 7×7 convolutions, firstly, the kernel size determines the receptive field size; a larger kernel results in more holistic feature acquisition. Secondly, camera noise is relatively weak, and when there are many categories, the distinction between classes may be concentrated on small local areas. Therefore, using this branch network alone is ineffective. However, MFCNN does not have these problems, so its accuracy does not decrease significantly with increasing device scale.

[0072] To address the issue of category expansion, retraining the source tracing model in large-scale experiments is extremely time-consuming; even retraining an MFCNN-based source tracing model requires at least 60 hours. This invention, however, utilizes high-dimensional vectors extracted from test images to form fingerprint indexes corresponding to the source camera category, and constructs a camera fingerprint database for retrieval and matching. This ensures source tracing accuracy while effectively reducing retrieval and matching time; the process of establishing the camera fingerprint database and retrieving the test image set can be completed in just four hours. Therefore, the above experiments demonstrate that this invention can achieve efficient device source tracing with minimal loss of accuracy.

[0073] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A source camera identification method based on multi-scale feature fusion, characterized in that, The method comprises the following steps: A traceability model is constructed based on multi-scale feature fusion, comprising a multi-scale feature extraction module, a feature fusion module and a classifier; the multi-scale feature extraction module comprises multiple feature extraction branches of different scales in parallel, each feature extraction branch processes the input image through convolution of the corresponding scale, and then sends the processed image to a residual network through a convolution layer, a batch normalization layer, an activation layer and a pooling layer to extract a deep vector; the residual network comprises a pooling layer, a first convolution block and two second convolution blocks connected in sequence, an activation layer and a pooling layer, a first residual connection is constructed between the output end of the pooling layer and the input end of the first second convolution block, the first residual connection passes through a convolution layer and a batch normalization layer, a second residual connection is constructed between the output end of the first convolution block and the input end of the second second convolution block, and a third residual connection is constructed between the output end of the first second convolution block and the input end of the activation layer and the pooling layer, wherein the first convolution block comprises a first size convolution layer, a batch normalization layer and an activation layer connected in sequence, a second size convolution layer, a batch normalization layer and an activation layer, and a first size convolution layer and a batch normalization layer, and the second convolution block comprises an activation layer and a pooling layer connected in sequence, a first size convolution layer, a batch normalization layer and an activation layer, a second size convolution layer, a batch normalization layer and an activation layer, and a first size convolution layer and a batch normalization layer; The traceability model is supervised learned by using a training image set, and a feature extractor is formed by extracting the multi-scale feature extraction module and the feature fusion module after supervised learning; A high-dimensional vector of each test image in a test image set is extracted by using the feature extractor, and the high-dimensional vector is used as a fingerprint index of a corresponding source camera class of the test image to construct a camera fingerprint library; A high-dimensional vector of a to-be-tested image is extracted by using the feature extractor, and a source camera class of the to-be-tested image is obtained based on matching retrieval of the high-dimensional vector and a fingerprint index in the camera fingerprint library. 2.The source camera identification method based on multi-scale feature fusion according to claim 1, characterized in that, The feature fusion module splices the deep vectors output by each feature extraction branch in the multi-scale feature extraction module to obtain a high-dimensional vector. 3.The source camera identification method based on multi-scale feature fusion according to claim 1, characterized in that, The classifier maps and activates the high-dimensional vector output by the feature fusion module by using a full connection layer and a Softmax layer to obtain a source camera recognition result. 4.The source camera identification method based on multi-scale feature fusion according to claim 1, characterized in that, The matching retrieval based on the high-dimensional vector and the fingerprint index in the camera fingerprint library comprises: Similarities of the high-dimensional vector and each fingerprint index in the camera fingerprint library are calculated, and a fingerprint index corresponding to a high similarity is selected as a matching retrieval fingerprint index.

5. The source camera identification method based on multi-scale feature fusion according to claim 4, characterized in that, The similarities of the high-dimensional vector and each fingerprint index in the camera fingerprint library are measured by using a Euclidean distance, a cosine distance or a Chebyshev distance.

6. A source camera identification device based on multi-scale feature fusion, implemented by using the source camera identification method based on multi-scale feature fusion in any one of claims 1-5, characterized in that, A model construction unit, a supervised learning unit, a camera fingerprint library construction unit and a matching retrieval unit are included; The model construction unit is configured to construct a traceability model based on multi-scale feature fusion, comprising a multi-scale feature extraction module, a feature fusion module and a classifier; The supervised learning unit is configured to supervise the learning of the traceability model by using a training image set, and to form a feature extractor by extracting the multi-scale feature extraction module and the feature fusion module after supervised learning; The camera fingerprint library construction unit is configured to extract a high-dimensional vector of each test image in the test image set by using the feature extractor, and construct a camera fingerprint library by using the high-dimensional vector as a fingerprint index of a corresponding source camera class of the test image; The matching retrieval unit is configured to extract a high-dimensional vector of the to-be-tested image by using the feature extractor, perform matching retrieval based on the high-dimensional vector and the fingerprint index in the camera fingerprint library, and obtain the source camera class of the to-be-tested image based on the fingerprint index of the matching retrieval.

7. A source camera identification device based on multi-scale feature fusion, comprising a memory and a processor, the memory is used to store a computer program, characterized in that, The processor is configured to implement the source camera identification method based on multi-scale feature fusion according to any one of claims 1-5 when executing the computer program.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is configured to implement the source camera identification method based on multi-scale feature fusion according to any one of claims 1-5 when the computer program is executed by the computer.

Images