An image tamper detection method based on multi-flow feature fusion

Abstract

The application discloses an image tampering detection method based on multi-flow feature fusion, relates to the fields of computer vision and information security, and specifically comprises the following steps: acquiring a tampered image dataset of a preset quantity and having labeled tampering type labels and tampering region positioning labels, and forming a sample set taking a single tampered image and corresponding tampering type labels and tampering region positioning labels as samples; using each sample in the sample set to train a multi-flow feature fusion network model; and using the trained multi-flow feature fusion network to test a to-be-tested image to obtain a final tampering detection result. The method can capture the diversity features of tampering operations from different dimensions, and effectively improves the accuracy and robustness of image tampering detection.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and information security, specifically to image tampering detection and digital image content verification. Background Technology

[0002] With the rapid development of image editing technology and the widespread availability of user-friendly software, image tampering has become easily accessible and low-cost, enabling precise and covert modification of images to achieve the purposes of forging evidence, concealing facts, and spreading false information. The authenticity and integrity of images are crucial for information security, digital forensics, and media verification; therefore, effectively detecting tampered areas in images has become an urgent problem to be solved in the field of image processing.

[0003] Currently, technologies for image tampering detection are mainly divided into traditional methods and deep learning-based methods. Traditional methods rely on manually designed features and rules, which are difficult to handle complex color and noise feature variations, making it difficult to accurately detect complex tampering scenarios.

[0004] Deep learning-based image tampering detection methods reduce reliance on hand-crafted features by automatically learning image features. However, existing deep learning-based image tampering detection methods focus on extracting single features, resulting in insufficient generalization. Furthermore, these methods struggle to accurately handle high-resolution images, exhibiting problems such as blurred detection boundaries and poor localization accuracy. In addition, deep learning-based detection techniques cannot effectively handle compression and noise. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an image tampering detection method based on multi-stream feature fusion.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for image tampering detection based on multi-stream feature fusion is proposed, which performs tampering detection on the input image according to steps S1 to S3 to obtain the tampering detection result of the input image:

[0008] Step S1: Obtain a preset number of tampered image datasets labeled with tamper type tags and tamper area location tags, and form a sample set with a single tampered image and its corresponding tamper type tag and tamper area location tag as samples;

[0009] Step S2: Construct a multi-stream feature fusion network that includes a preprocessing module, a multi-stream feature extraction module, a feature fusion module, and an output module. Use each sample in the sample set to train the multi-stream feature fusion network, which takes an image as input and outputs the corresponding tampering type and tampering region location map.

[0010] Step S3: Use the trained multi-stream feature fusion network to test the image under test to obtain the final tampering detection result.

[0011] Furthermore, the input of the preprocessing module constitutes the input of the multi-stream feature fusion network; the output of the preprocessing module is connected in series with the multi-stream feature extraction module, the feature fusion module, and the output module; the output of the output module constitutes the output of the multi-stream feature fusion network.

[0012] The preprocessing module receives the input image, extracts noise residual features from the input image to form a noise stream, and converts the input image into a YCrCb spatial image to form a YCrCb stream. It further converts each channel of the YCrCb spatial image into a frequency channel and stitches them together into a three-dimensional DCT cube to form a DCT stream. These three streams together constitute the output of the preprocessing module. The multi-stream feature extraction module extracts RoI features from the three streams in parallel. The feature fusion module fuses the RoI features from the three streams to obtain a weighted fusion feature of the input image. Based on the weighted fusion feature, the output module generates a classification result of the image tampering type and a tampering region localization map.

[0013] Furthermore, the preprocessing module includes a data augmentation module, a color space conversion module, a frequency domain transformation module, and an SRM filter kernel module, and the frequency domain transformation module includes a discrete cosine transform module and a selection splicing module;

[0014] The input terminal of the data augmentation module constitutes the input terminal of the preprocessing module; the input terminals of the color space conversion module and the SRM filter kernel module are respectively connected to the output terminal of the data augmentation module; the input terminal of the discrete cosine transform module is connected to one output terminal of the color space conversion module; the input terminal of the selection and splicing module is connected to the output terminal of the discrete cosine transform module; the output terminal of the preprocessing module consists of the output terminal of the SRM filter kernel module, the other output terminal of the color space conversion module, and the output terminal of the selection and splicing module.

[0015] The data augmentation module receives the input image, performs rotation and scaling operations to augment the image, and outputs the augmented image to the color space conversion module and the SRM filter kernel module. The color space conversion module converts the input image into a YCrCb space image using color space conversion formulas, forming a YCrCb stream. The discrete cosine transform module converts each channel of the YCrCb space image into frequency channels and outputs them to the selection and stitching module. The selection and stitching module normalizes the mean and variance of each frequency channel and stitches the processed frequency channels into a three-dimensional DCT cube, forming a DCT stream. The SRM filter kernel module extracts the noise residual features of the input image using an SRM filter, forming a noise stream.

[0016] Furthermore, the multi-stream feature extraction module includes a YCrCb convolutional feature extraction module, a DCT convolutional feature extraction module, a noisy convolutional feature extraction module, an RPN network module, and a RoI pooling module;

[0017] The input of the multi-stream feature extraction module consists of the inputs of the YCrCb convolutional feature extraction module, the DCT convolutional feature extraction module, and the noise convolutional feature extraction module; the input of the RPN network module is connected to the output of the YCrCb convolutional feature extraction module; the input of the RoI pooling module is connected to the outputs of the RPN network module, the DCT convolutional feature extraction module, and the noise convolutional feature extraction module; the output of the RoI pooling module constitutes the output of the multi-stream feature extraction module.

[0018] The YCrCb convolutional feature extraction module receives the YCrCb spatial image contained in the YCrCb stream, adjusts the weights of each channel of the YCrCb spatial image using a channel attention mechanism, and extracts deep spatial features of the YCrCb spatial image using a convolutional network, which is then output to the RPN network module. The RPN network module generates single-stream RoIs corresponding to the YCrCb stream and outputs them to the RoI pooling module. The DCT convolutional feature extraction module uses a dynamic gating mechanism to filter each frequency channel of the 3D DCT cube, and extracts deep frequency domain features from the filtered frequency channels using a convolutional network, which is then output to the RoI pooling module. The noise convolutional feature extraction module extracts deep noise features based on the noise residual features of the noise stream, using a spatial attention mechanism and a convolutional network, and outputs them to the RoI pooling module. The RoI pooling module generates RoI features corresponding to the YCrCb stream, the DCT stream, and the noise stream respectively through a matching and integration operation.

[0019] Furthermore, the feature fusion module includes a convolution module and a multi-level attention module;

[0020] The input of the convolution module constitutes the input of the feature fusion module; the input of the multi-level attention module is connected to the output of the convolution module; the output of the multi-level attention module constitutes the output of the feature fusion module.

[0021] The convolution module applies convolution operations to the RoI features corresponding to the YCrCb stream, the DCT stream, and the noisy stream, adjusting the dimensions of the RoI features of each stream to be consistent. Then, it performs a concatenation operation to generate a cascaded feature map, which is further output to the multi-level attention module. The multi-level attention module sequentially uses channel attention mechanism, dynamic gating mechanism, and spatial attention mechanism, combined with residual concatenation, to generate weight maps one by one. The weight maps are then multiplied element-wise with the cascaded feature maps to generate weighted fusion features.

[0022] Furthermore, the output module includes a bounding box regression prediction module, a classification module, and a localization map generation module;

[0023] The bounding box regression prediction module generates bounding box regression prediction maps of candidate regions based on the RoI features of the YCrCb stream and a preset confidence score threshold; the classification module processes the weighted fusion features with a fully connected layer to generate classification results for image tampering types; and the localization map generation module generates localization maps of tampered regions based on the weighted fusion features and through an attention weighting mechanism.

[0024] Furthermore, during model training of the multi-stream feature fusion network, joint optimization is performed using a total loss function that includes classification loss function, regression loss function, and RPN loss function. The classification loss function, regression loss function, RPN loss function, and total loss function are shown below:

[0025]

[0026] in, For the classification loss function, w i Let be the predicted probability of the i-th candidate bounding box. The foreground or background label corresponding to the i-th candidate bounding box;

[0027]

[0028] in, For the regression loss function, t i Predict coordinates for foreground candidate bounding boxes. For the coordinates of the input image bounding box, smooth L1 (x) represents the smoothed L1 loss;

[0029]

[0030] Among them, L RPN The RPN loss function is... For classifying losses, For the regression loss function, N C The number of candidate bounding boxes for classification loss, wi Let be the predicted probability of the i-th candidate bounding box. N represents the foreground or background label corresponding to the i-th candidate bounding box. R The number of candidate bounding boxes for regression loss, t i Predict coordinates for foreground candidate bounding boxes. The coordinates of the bounding box of the input image are given, and λ is the weight parameter for the classification loss and the regression loss.

[0031] L total =αL RPN +βL Loc

[0032]

[0033] Among them, L total Let L be the total loss function. RPN Let L be the RPN loss function. Loc Let be the localization loss function, α be the PRN loss weight coefficient, β be the localization loss weight coefficient, and M be the binary mask of the input image. This is the probability mask for the model's predictions.

[0034] The beneficial effects of adopting the above technical solution are as follows:

[0035] (1) This invention integrates the YCrCb color space, DCT frequency domain and noise flow features into a unified network. Through parallel processing and feature fusion, it can capture the diverse features of tampering operations from different dimensions, solving the problem of insufficient generalization in the prior art which focuses on the extraction of a single feature.

[0036] (2) This invention can capture tampering artifacts, edge artifacts and noise artifacts in images more accurately through YCrCb spatial transformation, DCT transformation and SRM filter kernel rotation operation, effectively improving the accuracy and robustness of image tampering detection, and solving the problems of blurred detection boundaries, poor positioning accuracy and inability to effectively handle compression and noise in the prior art. Attached Figure Description

[0037] Figure 1 This is a flowchart of the present invention;

[0038] Figure 2 This is a network architecture diagram of the multi-stream feature fusion network of the present invention;

[0039] Figure 3 This is a diagram of the preprocessing module architecture of the present invention;

[0040] Figure 4 This is a flowchart of the preprocessing process of the present invention;

[0041] Figure 5This is a diagram of the multi-level attention module architecture of the present invention;

[0042] Figure 6 The graph shows the comparison results of application examples of this invention. Detailed Implementation

[0043] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0044] refer to Figure 1 and Figure 2 A method for detecting image tampering based on multi-stream feature fusion is proposed, which performs detection on the input image according to steps S1 to S3 to obtain the tampering detection result of the input image:

[0045] Step S1: Obtain a preset number of tampered image datasets labeled with tamper type tags and tamper region location tags. The tamper types include splicing, removal, copying and moving, blurring, compression, adding artifacts, and overlapping. A sample set is formed with a single tampered image and its corresponding tamper type tag and tamper region location tag as samples.

[0046] Step S2: Construct a multi-stream feature fusion network that includes a preprocessing module, a multi-stream feature extraction module, a feature fusion module, and an output module. Use each sample in the sample set to train the multi-stream feature fusion network, which takes an image as input and outputs the corresponding tampering type and tampering region location map.

[0047] Step S3: Use the trained multi-stream feature fusion network to test the image under test to obtain the final tampering detection result.

[0048] Furthermore, the input of the preprocessing module constitutes the input of the multi-stream feature fusion network; the output of the preprocessing module is connected in series with the multi-stream feature extraction module, the feature fusion module, and the output module; the output of the output module constitutes the output of the multi-stream feature fusion network.

[0049] Specifically, the preprocessing module receives the input image, extracts noise residual features from the input image to form a noise stream, and converts the input image into a YCrCb spatial image to form a YCrCb stream. Furthermore, it converts each channel of the YCrCb spatial image into a frequency channel and stitches them together into a three-dimensional DCT cube to form a DCT stream. These three streams together constitute the output of the preprocessing module. The multi-stream feature extraction module extracts RoI features from the three streams in parallel. The feature fusion module fuses the RoI features from the three streams to obtain a weighted fusion feature of the input image. Based on the weighted fusion feature, the output module generates a classification result of the image tampering type and a tampering region localization map.

[0050] Furthermore, such as Figure 3As shown, the preprocessing module includes a data augmentation module, a color space conversion module, a frequency domain transformation module, and an SRM filter kernel module, and the frequency domain transformation module includes a discrete cosine transform module and a selection splicing module.

[0051] The input terminal of the data augmentation module constitutes the input terminal of the preprocessing module; the input terminals of the color space conversion module and the SRM filter kernel module are respectively connected to the output terminal of the data augmentation module; the input terminal of the discrete cosine transform module is connected to one output terminal of the color space conversion module; the input terminal of the selection and splicing module is connected to the output terminal of the discrete cosine transform module; the output terminal of the preprocessing module consists of the output terminal of the SRM filter kernel module, the other output terminal of the color space conversion module, and the output terminal of the selection and splicing module.

[0052] Specifically, such as Figure 4 As shown, the data augmentation module receives the input image, performs data augmentation on the input image using rotation and scaling operations, and outputs the data-augmented input image to the color space conversion module and the SRM filter kernel module. The color space conversion module uses the color space conversion formula to convert the input image into a YCrCb space image, forming a YCrCb stream. The discrete cosine transform module converts each channel of the YCrCb space image into frequency channels and outputs them to the selection and stitching module. The selection and stitching module performs mean and variance normalization on each frequency channel to reduce redundant channels, and stitches the processed frequency channels into a three-dimensional DCT cube, forming a DCT stream. The SRM filter kernel module uses the SRM filter to extract the noise residual features of the input image, forming a noise stream.

[0053] Furthermore, the multi-stream feature extraction module includes a YCrCb convolutional feature extraction module, a DCT convolutional feature extraction module, a noisy convolutional feature extraction module, an RPN network module, and a RoI pooling module;

[0054] The input of the multi-stream feature extraction module consists of the inputs of the YCrCb convolutional feature extraction module, the DCT convolutional feature extraction module, and the noise convolutional feature extraction module; the input of the RPN network module is connected to the output of the YCrCb convolutional feature extraction module; the input of the RoI pooling module is connected to the outputs of the RPN network module, the DCT convolutional feature extraction module, and the noise convolutional feature extraction module; the output of the RoI pooling module constitutes the output of the multi-stream feature extraction module.

[0055] Specifically, the YCrCb convolutional feature extraction module receives the YCrCb spatial image contained in the YCrCb stream, adjusts the weights of each channel of the YCrCb spatial image using a channel attention mechanism, and extracts deep spatial features of the YCrCb spatial image using a convolutional network, which is then output to the RPN network module. The RPN network module generates single-stream RoIs corresponding to the YCrCb stream and outputs them to the RoI pooling module. The DCT convolutional feature extraction module uses a dynamic gating mechanism to filter each frequency channel of the 3D DCT cube, and extracts deep frequency domain features from the filtered frequency channels using a convolutional network, which is then output to the RoI pooling module. The noise convolutional feature extraction module extracts deep noise features based on the noise residual features of the noise stream, using a spatial attention mechanism and a convolutional network, and outputs them to the RoI pooling module. The RoI pooling module generates RoI features corresponding to the YCrCb stream, the DCT stream, and the noise stream respectively through a matching and integration operation.

[0056] Furthermore, the feature fusion module includes a convolution module and a multi-level attention module;

[0057] The input of the convolution module constitutes the input of the feature fusion module; the input of the multi-level attention module is connected to the output of the convolution module; the output of the multi-level attention module constitutes the output of the feature fusion module.

[0058] Specifically, the convolution module applies convolution operations to the RoI features corresponding to the YCrCb stream, the DCT stream, and the noise stream, adjusting the RoI feature dimensions of each stream to be consistent, and then concatenating them to generate a cascaded feature map, which is further output to the multi-level attention module; for example, Figure 5 As shown, a multi-level attention module sequentially utilizes channel attention mechanism, dynamic gating mechanism, and spatial attention mechanism, combined with residual splicing, to generate a weight map. The weight map is then multiplied element-wise with the cascaded feature map to generate a weighted fusion feature.

[0059] Furthermore, the output module includes a bounding box regression prediction module, a classification module, and a localization map generation module; wherein, the bounding box regression prediction module generates a bounding box regression prediction map of the candidate region based on the RoI features of the YCrCb stream and combined with a preset confidence score threshold; the classification module is used to process the weighted fusion features through a fully connected layer to generate a classification result of the image tampering type; the localization map generation module generates a localization map of the tampered region based on the weighted fusion features and through an attention weighting mechanism.

[0060] Furthermore, during model training of the multi-stream feature fusion network, joint optimization is performed using a total loss function that includes classification loss function, regression loss function, and RPN loss function. The classification loss function, regression loss function, RPN loss function, and total loss function are shown below:

[0061] L C (w i ,w i * )=-[w i * log(w i )+(1-w i * log(1-w) i )]

[0062] in, For the classification loss function, w i Let be the predicted probability of the i-th candidate bounding box. The foreground or background label corresponding to the i-th candidate bounding box;

[0063]

[0064] in, For the regression loss function, t i Predict coordinates for foreground candidate bounding boxes. For the coordinates of the input image bounding box, smooth L1 (x) represents the smoothed L1 loss;

[0065]

[0066] Among them, L RPN The RPN loss function is... For classifying losses, For the regression loss function, N C The number of candidate bounding boxes for classification loss, w i Let be the predicted probability of the i-th candidate bounding box. N represents the foreground or background label corresponding to the i-th candidate bounding box. R The number of candidate bounding boxes for regression loss, t i Predict coordinates for foreground candidate bounding boxes. The coordinates of the bounding box of the input image are given, and λ is the weight parameter for the classification loss and the regression loss.

[0067] L total =αL RPN +βL Loc

[0068]

[0069] Among them, L total Let L be the total loss function. RPN Let L be the RPN loss function. Loc Let be the localization loss function, α be the PRN loss weight coefficient, β be the localization loss weight coefficient, and M be the binary mask of the input image. This is the probability mask for the model's predictions.

[0070] Furthermore, to verify the effectiveness of the present invention, it was compared with H-LSTM, HP-FCN, Mantra-Net, SPAN, MVSS-Net, MVSS-Net++, and OSN based on the CASIAv1 and NIST16 datasets. Specifically, the model was trained using the CASIAv2 dataset containing 5123 images, including 3295 copy-paste tampering samples and 1828 spliced ​​tampering samples. The model was then tested using the CASIAv1 and NIST16 datasets. The F1 score with a fixed threshold was used to measure the detection accuracy of each model in the tampered region. The test comparison results are shown in Table 1 and the test comparison results are illustrated in the figure below. Figure 6 As shown.

[0071] Table 1

[0072]

[0073] As shown in Table 1, the present invention outperforms H-LSTM, HP-FCN, ManTra-Net, SPAN, MVSS-Net, MVSS-Net++, and OSN in F1 scores on the CASIAVL, NIST16, and MEAN datasets; Figure 6 As shown, by comparing the present invention with ManTra-Net, MVSS-Net, and PSCC-Net, the accuracy and clarity of the present invention are qualitatively demonstrated. Furthermore, the present invention provides the closest result for tampering region localization to the ground truth, i.e., the tampering region localization map corresponding to each tampered image.

[0074] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A multi-stream feature fusion-based image tamper detection method, characterized in that, Perform steps S1 to S3 to detect tampering in the input image and obtain the tampering detection results. Step S1: Obtain a preset number of tampered image datasets labeled with tamper type tags and tamper area location tags, and form a sample set with a single tampered image and its corresponding tamper type tag and tamper area location tag as samples; Step S2: Construct a multi-stream feature fusion network that includes a preprocessing module, a multi-stream feature extraction module, a feature fusion module, and an output module. Use each sample in the sample set to train the multi-stream feature fusion network, which takes an image as input and outputs the corresponding tampering type and tampering region location map. Step S3: Use the trained multi-stream feature fusion network to test the image under test to obtain the final tamper detection result; Specifically, the preprocessing module includes a data augmentation module, a color space conversion module, a frequency domain transformation module, and an SRM filter kernel module, and the frequency domain transformation module includes a discrete cosine transform module and a selection splicing module; The input terminal of the data augmentation module constitutes the input terminal of the preprocessing module; the input terminals of the color space conversion module and the SRM filter kernel module are respectively connected to the output terminal of the data augmentation module; the input terminal of the discrete cosine transform module is connected to one output terminal of the color space conversion module; the input terminal of the selection and splicing module is connected to the output terminal of the discrete cosine transform module; the output terminal of the preprocessing module consists of the output terminal of the SRM filter kernel module, the other output terminal of the color space conversion module, and the output terminal of the selection and splicing module. The data augmentation module receives the input image, performs data augmentation on the input image using rotation and scaling operations, and outputs the data-enhanced input image to the color space conversion module and the SRM filter kernel module. The color space conversion module converts the input image into a YCrCb space image using color space conversion formulas, forming a YCrCb stream. The discrete cosine transform module converts each channel of the YCrCb space image into a frequency channel and outputs it to the selection and stitching module. The selection and stitching module normalizes the mean and variance of each frequency channel and stitches the processed frequency channels into a three-dimensional DCT cube, forming a DCT stream. The SRM filter kernel module uses an SRM filter to extract the noise residual features of the input image, forming a noise stream. The multi-stream feature extraction module includes a YCrCb convolutional feature extraction module, a DCT convolutional feature extraction module, a noisy convolutional feature extraction module, an RPN network module, and a RoI pooling module; The input of the multi-stream feature extraction module consists of the inputs of the YCrCb convolutional feature extraction module, the DCT convolutional feature extraction module, and the noise convolutional feature extraction module; the input of the RPN network module is connected to the output of the YCrCb convolutional feature extraction module; the input of the RoI pooling module is connected to the outputs of the RPN network module, the DCT convolutional feature extraction module, and the noise convolutional feature extraction module; the output of the RoI pooling module constitutes the output of the multi-stream feature extraction module. The YCrCb convolutional feature extraction module receives the YCrCb spatial image contained in the YCrCb stream, adjusts the weights of each channel of the YCrCb spatial image using a channel attention mechanism, and extracts the deep spatial features of the YCrCb spatial image using a convolutional network, which is then output to the RPN network module. The RPN network module generates single-stream RoIs corresponding to the YCrCb stream and outputs them to the RoI pooling module. The DCT convolutional feature extraction module uses a dynamic gating mechanism to filter each frequency channel of the 3D DCT cube, and extracts the depth frequency domain features from the filtered frequency channels using a convolutional network, which is then output to the RoI pooling module. The noise convolution feature extraction module extracts deep noise features based on the noise residual features of the noise flow, utilizes spatial attention mechanism and convolutional network, and outputs them to the RoI pooling module. The RoI pooling module generates RoI features corresponding to the YCrCb stream, the DCT stream, and the noise stream respectively through a matching and integration operation.

2. The image tampering detection method based on multi-flow feature fusion according to claim 1, characterized in that, The input of the preprocessing module constitutes the input of the multi-stream feature fusion network; the output of the preprocessing module is connected in series with the multi-stream feature extraction module, feature fusion module, and output module; the output of the output module constitutes the output of the multi-stream feature fusion network. The preprocessing module receives the input image, extracts noise residual features from the input image to form a noise stream, and converts the input image into a YCrCb spatial image to form a YCrCb stream. It further converts each channel of the YCrCb spatial image into a frequency channel and stitches them together into a three-dimensional DCT cube to form a DCT stream. These three streams together constitute the output of the preprocessing module. The multi-stream feature extraction module extracts the RoI features of the three streams in parallel. The feature fusion module fuses the RoI features of the three streams to obtain the weighted fused features of the input image. The output module generates classification results of image tampering types and tampering region location maps based on weighted fusion features. 3.The method of claim 2, wherein, The feature fusion module includes a convolution module and a multi-level attention module; The input of the convolution module constitutes the input of the feature fusion module; the input of the multi-level attention module is connected to the output of the convolution module; the output of the multi-level attention module constitutes the output of the feature fusion module. The convolution module applies convolution operations to the RoI features corresponding to the YCrCb stream, the DCT stream, and the noise stream, adjusting the dimensions of the RoI features of each stream to be consistent, and then performs a concatenation operation to generate a cascaded feature map, which is further output to the multi-level attention module. The multi-level attention module sequentially utilizes channel attention mechanism, dynamic gating mechanism, and spatial attention mechanism, combined with residual splicing, to generate weight maps one by one. The weight maps are then multiplied element-wise with the cascaded feature maps to generate weighted fusion features.

4. The image tamper detection method based on multi-stream feature fusion according to claim 2, characterized in that, The output module includes a bounding box regression prediction module, a classification module, and a localization map generation module; The bounding box regression prediction module generates bounding box regression prediction maps of candidate regions based on the RoI features of the YCrCb stream and a preset confidence score threshold; the classification module processes the weighted fusion features with a fully connected layer to generate classification results for image tampering types; and the localization map generation module generates localization maps of tampered regions based on the weighted fusion features and through an attention weighting mechanism.

5. The image tamper detection method based on multi-stream feature fusion according to claim 1, characterized in that, When training the multi-stream feature fusion network, a joint optimization is performed using a total loss function that includes the classification loss function, regression loss function, and RPN loss function. The classification loss function, regression loss function, RPN loss function, and total loss function are shown below: ； wherein, is a classification loss function, is a prediction probability of the th candidate bounding box, is a foreground label or a background label corresponding to the th candidate bounding box; ； ； wherein, is a regression loss function, is a foreground candidate bounding box prediction coordinate, is an input image bounding box coordinate, is a smooth L1 loss; ； in, The RPN loss function is... For classifying losses, For regression loss function, The number of candidate bounding boxes is used for classification loss. For the first The predicted probability of each candidate bounding box. For the first Foreground or background labels corresponding to each candidate bounding box To regress the number of candidate bounding boxes, Predict coordinates for foreground candidate bounding boxes. The input image bounding box coordinates, These are the weighting parameters for classification loss and regression loss; ； ； wherein, is a total loss function, is a RPN loss function, is a localization loss function, is a PRN loss weight coefficient, is a localization loss weight coefficient, is a binary mask of the input image, is a probability mask predicted by the model.

Images