Attention mechanism driven digital image tampering detection system

Abstract

The invention discloses a digital image tampering detection system driven by an attention mechanism, and belongs to the technical field of image security authentication. In order to solve the problems that an existing method is low in sensitivity to complex tampering traces and insufficient in positioning precision, the system firstly carries out standardized preprocessing on an input image; then multi-scale features are extracted through a deep convolutional neural network; a double-path attention mechanism is innovatively introduced, wherein a space attention module focuses on the space position of a tampering area, and a channel attention module adaptively enhances feature channel response related to tampering operation; the fused and optimized attention features are subjected to up-sampling through a decoder to recover the resolution, and finally a pixel-level tampering positioning mask and image-level tampering probability judgment are synchronously output; according to the method, the detection robustness of fine tampering traces is remarkably improved through an attention-driven mechanism, and an efficient solution is provided for authenticity identification of digital images while the false alarm rate is reduced.

Description

Technical Field

[0001] This invention belongs to the field of digital image security authentication and anti-counterfeiting technology, specifically relating to an attention mechanism-driven digital image tampering detection system. Background Technology

[0002] As a core carrier of information dissemination, the authenticity of digital images directly affects the credibility of news, the validity of judicial evidence, and the security of digital assets. With the intelligent development of editing tools, the concealment of tampering operations (such as copy-paste, splicing, and generative content replacement) has significantly increased, necessitating high-precision detection technologies to ensure image credibility. Deep learning methods, represented by convolutional neural networks, have demonstrated powerful advantages in automatic feature extraction and have been widely applied in the field of image tampering detection. By learning common patterns of tampering traces (such as edge discontinuities and abnormal noise distribution) on datasets such as CASIA and COVERAGE, they significantly surpass the detection accuracy of traditional methods based on manual features (such as JPEG block artifact analysis), providing a new paradigm for forgery identification.

[0003] However, existing technologies suffer from fundamental flaws: mainstream models process feature responses across all image regions equally, failing to actively focus on tamper-sensitive areas (such as minute boundary artifacts and local texture conflicts), resulting in persistently high false positive rates under complex background interference. Simultaneously, the local receptive field of standard convolutional operations struggles to capture long-distance dependencies (such as the correlation between the source and target regions during copy-and-move operations), and its generalization ability to post-processed (blurring, noise injection) tampering traces is insufficient. With the upgrading of adversarial tampering techniques, these problems will severely restrict the practical application of detection systems, necessitating the development of a next-generation detection architecture with adaptive feature optimization capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide an attention mechanism-driven digital image tampering detection system. It addresses three major shortcomings of existing detection technologies: insufficient focus on tamper-sensitive areas due to equal processing of image features, weak generalization ability to post-processing operations (such as blurring and noise injection), and difficulty in modeling long-distance tampering correlations (such as cross-regional consistency of copy-movement). By introducing an adaptive attention-driven mechanism, the system significantly improves the accuracy and robustness of tampering trace localization.

[0005] The technical solution of this invention is: an attention mechanism-driven digital image tampering detection system, the core process of which includes: Preferably, the input image is first preprocessed using standardization, such as size normalization and pixel value normalization. Then, a deep convolutional neural network, such as ResNet-50, is used to extract multi-scale feature maps to capture tampering clues from low-level texture to high-level semantics. The key innovation lies in the design of a dual-path attention optimization module: the spatial attention branch compresses through channel pooling, such as concatenating max pooling and average pooling, to generate a high-resolution heatmap that accurately focuses on microscopic abnormal areas such as edge breaks and local texture conflicts; the channel attention branch uses a fully connected network with a bottleneck layer to learn feature channel weights and adaptively enhances feature responses that are strongly correlated with tampering operations such as copy-paste artifacts and splicing color differences. After fusing the two types of attention weights to modulate the original features, they are input into the U-Net architecture decoder for upsampling and skip connections, and finally output a pixel-level binary tampering mask map (accurate to pixel-level tampering region localization) and an image-level tampering probability score (overall authenticity determination).

[0006] As a further improvement of this invention, to address the spatial sparsity (only local areas are modified) and weakness of tampering traces, such as the concealment of traces due to post-processing masking, this invention proposes a dual-path attention collaborative mechanism: 1. Spatial attention introduces dilated convolution to enhance the receptive field, capturing large-span tampering associations while maintaining feature map resolution, such as matching source and target regions that have been copied and moved. 2. Channel attention combined with frequency domain decomposition technology separates the high-frequency residual components (sensitive to edge anomalies) and low-frequency structural components of the image, and specifically enhances high-frequency tampering features; 3. In the decoding stage, multi-task loss joint optimization is adopted. Pixel-level cross-entropy loss constrains the positioning accuracy, and image-level adversarial loss improves the ability to detect generative tampering, such as DeepFake.

[0007] Validated on the COCO-Tamper and CASIA v2 datasets, compared with traditional methods (such as ELA, NOI1) and mainstream deep learning models (MantraNet, RGB-N), this system improves the mIoU of detection for three types of tampering: copy-paste, splicing, and removal by 12.7%, improves the robustness of post-processing for Gaussian blur (σ=1.5) and JPEG compression (QF=70) by about 18.3%, and reduces the false alarm rate to 5.1%.

[0008] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention has high positioning accuracy: spatial attention is focused on tamper-sensitive areas, such as boundary transition zones, making faint traces visible; 2. This invention has strong generalization ability: the channel attention dynamic enhancement anti-interference feature, such as high-frequency residual, effectively resists post-processing distortion; 3. This invention optimizes efficiency through a lightweight dual-path attention design, increasing computational overhead by less than 8%, and achieving a single-graph inference time of ≤0.3 seconds on a 1080Ti GPU, meeting real-time detection requirements. Experiments show that, across datasets, migrating from CASIA to IMD2020, it maintains an average accuracy of 87.6%, significantly outperforming traditional models. Attached Figure Description

[0009] Figure 1 Image preprocessing diagram; Figure 2 For feature extraction; Figure 3 To tamper with the detection and analysis chart; Figure 4 Optimize the visualization results of the algorithm; Figure 5 For performance evaluation; Figure 6 Error report image; Figure 7a For test cases Figure 1 ; Figure 7b For test cases Figure 2 ; Figure 7c For test cases Figure 3 . Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0011] This embodiment proposes an attention mechanism-driven digital image tampering detection system, such as... Figure 1 As shown, the specific steps are as follows: Step 1: Image preprocessing: Perform standardization operations such as normalization and resizing on the digital image to be detected; Step S1a: Size standardization: The input image is scaled to a fixed resolution (e.g., 512×512 pixels), and a bilinear interpolation algorithm is used to maintain the integrity of the geometric structure. If the aspect ratio of the original image is too different, an edge filling strategy (filling with a gray value of 128) is used to maintain the content proportion and avoid distortion.

[0012] Step S1b: Pixel normalization: Perform a linear transformation of pixel values: Inorm=

[0013] Mapping pixel values ​​from the original range [0, 255] to [-1, 1] accelerates model convergence and improves numerical stability.

[0014] Step S1c: Color space conversion: For specific types of tampering, such as Deepfake skin color spoofing, the RGB image is converted to the YCbCr color space, separating the luminance component (Y) and chrominance component (Cb / Cr) to enhance sensitivity to color anomalies.

[0015] Step 2: Deep Feature Extraction: Extract multi-level feature maps from the preprocessed image using a convolutional neural network (CNN); Step S2a: Backbone network initialization: Load a pre-trained ResNet-50 model (trained on ImageNet), remove its fully connected classification layers, and retain the convolutional layers as feature encoders. Freeze the weights of the first two convolutional layers (Conv1, Conv2_x) to retain general edge / texture extraction capabilities, and fine-tune the deep convolutional layers (Conv3_x to Conv5_x) to adapt to the modified feature learning.

[0016] Step S2b: Multi-scale feature map generation: Input a preprocessed image and extract feature maps at four levels: C1:Conv2_x output layer ($H / 4\timesW / 4\times256$), preserving high-resolution details (such as sharp edges and subtle textures); C2: Conv3_x output layer ($H / 8 times W / 8 times 512$), captures intermediate semantics (such as local structural anomalies). C3: Conv4_x output layer ($H / 16\timesW / 16\times1024$), models the relationships between tampered regions; C4:Conv5_x output layer ($H / 32\timesW / 32\times2048$), encodes global context information.

[0017] Step S2c: Feature Pyramid Enhancement: Introducing a Feature Pyramid Network (FPN) to fuse multi-scale features through top-down paths and lateral connections: 1. Perform a 1×1 convolution on C4 to reduce the dimensionality to 256 channels; 2. Upsample sequentially and add the samples to the features of the previous level (C3→C2→C1); 3. Output enhanced feature maps {P1, P2, P3, P4} (resolution same as C1-C4, number of channels uniformly 256) to strengthen the consistency of cross-scale tampering clues.

[0018] Step 3: Spatial Attention Focusing: Apply the spatial attention mechanism to calculate the weight of each spatial location in the feature map, focus on regions that may contain traces of tampering (such as discontinuous edges or abnormal textures), and suppress irrelevant backgrounds; Step S3a: Feature channel compression: Perform bipath aggregation along the channel dimension on the input feature map F∈RH×W×C: Max pooling: Fmax = MaxPool(F) ∈ RH × W × 1.

[0019] Average pooling: Favg = AvgPool(F) ∈ RH × W × 1; Step S3b: Generating a space heat map: Spatial weights are learned through convolutional layers: As=σ(Conv7×7(Conv3×3(Fcat)), Wherein: Conv3×3: 3×3 convolution (number of channels 64→32), ReLU activation.

[0020] Conv7×7: 7×7 convolution (32 channels → 1), Sigmoid activation (σ).

[0021] Step S3c: Feature modulation and background suppression: Apply spatial weights to the original feature map: Fatt_sp = F⊗Expand(As); Where ⊗ represents element-wise multiplication; Expand(As) copies As C times along the channel dimension to RH×W×C.

[0022] Step 4: Channel Attention Enhancement: Apply the channel attention mechanism to calculate the importance weights of different feature channels, and adaptively enhance the response of feature channels that are highly correlated with the tampering type (such as copy and paste, splicing); Step S4a: Global feature descriptor generation: Perform channel-level statistics on the input feature map F∈RH×W×C: Global average pooling: zavgc= ; Global max pooling: zmaxc = maxi, j( ); Step S4b: Channel weight learning: Weights are learned through a fully connected network with a bottleneck layer: Ac = σ(W2·δ(W1·z+b1)+b2).

[0023] Where: W1∈RC / r×2C, W2∈RC×C / r (r=16 is the compression ratio); δ is the ReLU activation function, and σ is the Sigmoid function.

[0024] The output channel weight vector Ac∈RC×1 (value range [0,1]) corresponds to tamper-sensitive features (such as high-frequency residuals and chromaticity anomalies).

[0025] Step S4c: Feature channel modulation: Apply channel weights to the original feature map: Fatt_ch=F⊗Ac; Where ⊗ represents channel broadcast multiplication: all elements of the c-th channel of the feature map are multiplied by Ac(c).

[0026] Step 5: Attention Feature Fusion and Optimization: The weights of spatial attention and channel attention are fused to perform weighted optimization on the original feature map, highlighting key tampering clues; Step S5a: Attention weight fusion: Integrate spatial attention weights As∈RH×W×1 and channel attention weights Ac∈R1×1×C: Afusion=α⋅Expand(As)+β⋅Broadcast(Ac); Where: α=0.6, β=0.4 (optimized by grid search); Expand(As): Copy to RH×W×C; Broadcast(Ac): Broadcast to RH×W×C.

[0027] Step S5b: Feature Adaptive Weighting: Modulate the original feature map F: Fopt = F⊙exp(Afusion); Where ⊙ represents element-wise multiplication, and the exponential operation exp(⋅) enhances non-linearity: When Afusion > 0.8, feature amplification > 2.2 times (enhancing key clues); When Afusion < 0.3, the feature decays to < 74%.

[0028] Step S5c: Residual Link Optimization: Introducing cross-module residual learning to preserve original effective features: Ffinal=γ⋅Fopt+(1-γ)⋅F; Dynamic learning of parameter γ: γ = Sigmoid(Conv1×1(Concat(F, Fopt))), Output optimized feature Ffinal∈RH×W×C; Step 6: Feature Decoding and Upsampling: Input the optimized attention features into the decoder (such as a U-Net structure), and gradually restore the spatial resolution through upsampling and skip connections; Step S6a: Multi-level feature fusion: Construct a U-Net symmetric decoding path and fuse encoder features (C1-C4) and attention optimization features (Ffinal) layer by layer. Skip connection: The encoder features (such as C3∈ℝH / 8×W / 8×512) are concatenated with the same-scale decoder features (D3∈ℝH / 8×W / 8×256) channels, and then unified to 256 channels by 1×1 convolution.

[0029] Step S6b: Progressive upsampling: A hybrid strategy of transposed convolution and bilinear interpolation is adopted: 1. Transposed Convolution (Deconv): Core size 4×4, step size 2, fill size 1. The number of passages per floor is halved (256→128→64→32); 2. Bilinear interpolation: Interpolation thinning is performed after transpose convolution to eliminate chessboard artifacts.

[0030] Step S6c: Resolution Restoration and Output Final decoding process: D4 = Conv 3×3 (Ffinal); D3 = Conv 3×3 (Concat(C3, Deconv(D4))); D2 = Conv 3×3 (Concat(C2, Deconv(D3))); D1 = Conv 3×3 (Concat(C1, Deconv(D2))); Output of the tampered heatmap: Mask = Sigmoid(Conv 1×1 (D1)); Step 7: Tampering Region Location and Judgment: Output a pixel-level tampering region segmentation mask to accurately locate the tampered pixels, and output an image-level tampering probability to determine whether the entire image has been tampered with.

[0031] In recent years, deep learning technology has made significant progress in the field of image processing, especially the application of attention mechanisms, which enables models to focus more on key information in images, improving the accuracy and efficiency of detection.

[0032] This project aims to combine attention mechanisms and deep learning algorithms to develop an efficient and accurate digital image tampering detection system to address the increasingly serious problem of image tampering and protect the authenticity and integrity of information.

[0033] The system will provide a user-friendly interface and support various image processing and analysis functions, such as image resizing, cropping, and filtering. It will also combine image feature extraction, performance comparison, and data display modules to provide users with a comprehensive image tampering detection solution.

[0034] like Figure 1 As shown, the image preprocessing stage of this system mainly includes the following stages: 1. Image path input: Users can input the path to the image file to be processed.

[0035] 2. Image resizing: Users can input the desired image width and height to resize the image.

[0036] 3. Image cropping: Users can specify the X and Y coordinates, width, and height for cropping to perform image cropping operations.

[0037] 4. Filtering: Users can select different filter types (mean filter, median filter, Gaussian filter) to process images.

[0038] 5. Processing Results Display: Displays the results of image resizing, cropping, and filtering.

[0039] 6. Image Preview: Displays a comparison of pixel values ​​between the original and processed images in chart form.

[0040] like Figure 2 The image shows the feature extraction stage of the system: 1. Image Upload: Users can upload digital images from which features need to be extracted.

[0041] 2. Feature type selection: Users can select the feature types to be extracted, including feature type 1, feature type 2 and feature type 3.

[0042] 3. Parameter settings: Users can input or adjust the parameters required for feature extraction.

[0043] 4. Feature Extraction Control: Users can control the feature extraction process by clicking the "Start Extraction" and "Stop Extraction" buttons.

[0044] 5. Extraction Results Display: Shows the extracted feature names and their corresponding feature values.

[0045] 6. Eigenvalue distribution chart: Use a bar chart to show the distribution of eigenvalues.

[0046] 7. Eigenvalue Trend Chart: Use a line chart to show the trend of eigenvalue changes.

[0047] 8. Feature Correlation Chart: Use heatmaps to show the correlation between different features.

[0048] like Figure 3 The diagram shown is a tamper detection and analysis diagram of the system, mainly divided into two parts: user operation area: It provides an image upload function, allowing users to upload digital images to be detected.

[0049] Users can select the analysis type, including two modes: quick analysis and in-depth analysis.

[0050] Users can enter descriptions of the images.

[0051] A button to start analysis is provided; once the user clicks it, the system will begin analyzing the uploaded image.

[0052] Data display area: A pie chart showing the tampering detection results displays the distribution of normal, tampered, and uncertain results.

[0053] A bar chart showing the distribution of tampering types, illustrating the quantity distribution of different tampering types.

[0054] The detailed test items and results are presented in tabular form, including information such as integrity, degree of tampering, and type of tampering.

[0055] like Figure 4 The diagram shown represents the algorithm optimization results of this system, which mainly consists of the following four parts: 1. Algorithm Parameter Settings: Users can select the optimizer type (such as Adam, SGD, RMSprop), input the learning rate, and batch size in this area to configure the algorithm parameters.

[0056] 2. Start Optimization Button: When the user clicks this button, the system will start optimizing the algorithm according to the set parameters.

[0057] 3. Optimization Log Display: This area displays the number of iterations, loss value, and accuracy during the algorithm optimization process, allowing users to understand the optimization progress in real time.

[0058] 4. Visualization of the optimization process: Through two charts (changes in loss value and change in accuracy), users can intuitively observe the trends in the change of loss value and accuracy during the algorithm optimization process.

[0059] like Figure 5 The diagram shown is a performance evaluation chart for this system, which mainly consists of the following four parts: 1. Performance Parameter Input: Users can input the model's name, precision, recall, and F1 score to submit the model's performance data.

[0060] 2. Performance Comparison: Users can select two or more models from the drop-down menu to compare and analyze their performance differences.

[0061] 3. Performance data display: The precision, recall, and F1 score of each model are displayed in chart form so that users can intuitively understand the model performance.

[0062] 4. Detailed data: The performance data of each model is displayed in tabular form, including model name, precision, recall and F1 score, so that users can view detailed information.

[0063] like Figure 6 The diagram shown is a diagram of the error reporting system, which mainly consists of the following eight steps: 1. Error type selection: Users can select the type of error here, such as type 1, type 2 or type 3.

[0064] 2. Error Description Input: Users can enter a detailed description of the error here to facilitate more accurate recording and feedback of the problem.

[0065] 3. Error Count Input: Users need to input the number of times errors occurred in order to statistically analyze the frequency of errors.

[0066] 4. Error impact range input: Users need to input the impact range of the error, such as small range, medium range or large range.

[0067] 5. Error Date Input: Users need to input the date the error occurred to facilitate tracking and analyzing the time pattern of error occurrences.

[0068] 6. Submit an error report: After completing the error information, the user can click this button to submit an error report.

[0069] 7. Error Type Statistics Chart Display: The system automatically generates statistical charts of error types, displaying the distribution of various error types in the form of pie charts.

[0070] 8. Error Log Display: The system displays all submitted error logs in tabular form, including error type, error description, number of occurrences, scope of impact, and error date.

[0071] Figures 7(a)(b)(c) show the test case diagrams of this system. It can be seen that this system significantly enhances the detection capability of subtle tampering (<32×32 pixels) through the dual-path attention mechanism (spatial + channel), and maintains high robustness against post-processing operations such as blurring and compression. At the same time, the performance degradation is minimal in cross-dataset tests, and real-time detection of 35FPS can be achieved on a 3070Ti GPU, providing a high-precision and high-efficiency solution for scenarios such as news authenticity verification and judicial evidence collection.

[0072] The proposed attention-based digital image tampering detection system offers several advantages and benefits, primarily reflected in the following aspects: 1. High-precision detection capability: Through an innovative dual-path attention mechanism (spatial + channel), it achieves very high mIoU detection accuracy; the detection rate of subtle tampering (<32×32 pixels) is higher, which is a significant improvement over traditional methods.

[0073] 2. Strong anti-interference performance: It maintains high detection accuracy for post-processing operations such as Gaussian blur (σ=2.0) and JPEG compression (QF=70); and controls the false alarm rate to within 5.1% under complex background interference.

[0074] 3. Excellent generalization ability: It achieves higher accuracy in detecting novel generative tampering (such as DeepFake, StyleGAN) and adapts to different tampering types (copy-paste, splicing, content removal, etc.).

[0075] 4. High-efficiency real-time processing: Achieves a processing speed of 35FPS (512×512 resolution) on a 3070Ti GPU, with 4K ultra-high-definition image processing time controlled within 300ms, and the number of model parameters is only 8% higher than the baseline U-Net, demonstrating excellent computational efficiency.

[0076] 5. Wide range of application value: It provides reliable technical support for verifying the authenticity of news, has important application value in the field of judicial electronic evidence authentication, effectively serves digital copyright protection and content security supervision, and is particularly suitable for the detection needs of AI-generated content.

[0077] 6. Technological Innovation: It pioneered a spatial-channel dual-path attention collaboration mechanism, proposed a dynamic feature modulation strategy to achieve precise focusing on tampered regions, designed a lightweight architecture to balance detection accuracy and computational efficiency, developed a multi-task joint training framework, and simultaneously optimized localization and classification performance.

[0078] Through the aforementioned advantages, this patented technical solution has achieved a major breakthrough in detection accuracy, robustness, and practicality in the field of digital image authenticity authentication, providing an effective solution to address the increasingly complex challenges of image tampering.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

[0080] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An attention mechanism driven digital image tampering detection system, characterized in that, The method comprises the following steps: Step 1: image preprocessing: normalizing, resizing and other standardization operations are performed on the digital image to be detected; Step 2: deep feature extraction: a convolutional neural network is used to extract multi-level feature maps from the preprocessed image; Step 3: spatial attention focusing: a spatial attention mechanism is applied to calculate the weight of each spatial position of the feature map, focusing on the area that may contain tampering traces, and suppressing irrelevant background; Step 4: channel attention enhancement: a channel attention mechanism is applied to calculate the importance weight of different feature channels, and the response of the feature channel sensitive to a specific tampering type is adaptively enhanced; Step 5: attention feature fusion and optimization: the spatial attention and channel attention weights are fused, and the original feature map is weighted and optimized to highlight the key tampering clues; Step 6: feature decoding and upsampling: the optimized attention features are input into the decoder to gradually restore the spatial resolution through upsampling and skip connection; Step 7: tampering area positioning and determination: output the pixel-level tampering area segmentation mask map, and output the image-level tampering probability.

2. The attention mechanism driven digital image tampering detection system of claim 1, wherein: In step 1, the input digital image to be detected is standardized, mainly including uniform adjustment of image size to a preset resolution of the model, and normalization of pixel values to a specific range. In the model training stage, various data enhancement techniques are applied, including but not limited to random rotation, horizontal or vertical flipping, adding Gaussian noise or simulating JPEG compression artifacts. 3.The attention mechanism driven digital image tampering detection system of claim 1, wherein: In step 2, a pre-trained deep convolutional neural network is used as the backbone network to process the preprocessed image. The backbone network extracts low-level to high-level semantic features of the image layer by layer, generating a set of feature maps with different spatial resolutions and semantic abstraction levels. Multi-scale features jointly constitute the basis for subsequent analysis, where shallow features retain detailed information and deep features encode more abstract context information.

4. The attention mechanism driven digital image tampering detection system of claim 1, wherein: In step 3, the spatial attention mechanism acts on the extracted feature map. By calculating the importance weight of each position in the spatial dimension of the feature map, based on the information after channel pooling, the mean or maximum value of each position in all channels is calculated, and then a heat map is generated through a convolution layer. The heat map highlights the areas in the image that may contain tampering traces, suppresses irrelevant or low-information background areas, and guides the model to focus computing resources on the most suspicious spatial positions.

5. The attention mechanism driven digital image tampering detection system of claim 1, wherein: In step 4, the channel attention mechanism analyzes the importance of different feature channels. First, global pooling is used to compress the spatial information of each channel to obtain a channel description vector. Then, a small neural network is used to learn the weight of each channel. This weight adaptively enhances the response of the feature channel sensitive to a specific tampering type, and suppresses irrelevant or redundant channels, thereby focusing on the most discriminative tampering clues in the feature dimension.

6. The attention mechanism driven digital image tampering detection system of claim 1, wherein: In step 5, the spatial weight heat map output by the spatial attention module and the channel weight vector output by the channel attention module are fused. The original feature map is weighted and modulated using the fused attention weight, i.e., each spatial position and each channel element in the original feature map is multiplied by the corresponding attention weight.

7. The attention mechanism driven digital image tampering detection system of claim 1, wherein: In step 6, the optimized attention features are input into the decoder, which gradually recovers the spatial resolution of the feature map through upsampling operations. The decoder fuses shallow features from the corresponding levels of the encoder through a skip connection to supplement the details lost during the encoding process. After multiple levels of upsampling and feature fusion, a high-resolution feature map is finally generated, preparing for pixel-level positioning.

8. The attention mechanism driven digital image tampering detection system of claim 1, wherein: In step 7, this is the output layer of the system. The decoded high-resolution feature map is passed through a 1x1 convolution layer to generate a binary classification probability value for each pixel, forming a pixel-level tampered region segmentation mask image that accurately outlines the tampered pixel region. At the same time, by performing global pooling on the final feature map and connecting a fully connected layer, an image-level tampering probability value between 0 and 1 is output, which is used for overall judgment of whether the image has been tampered with. The system finally provides both pixel-level accurate positioning and image-level rapid classification results.

Images