Image tampering positioning method and system based on three-dimensional reconstruction technology

Abstract

The invention discloses an image tampering positioning method and system based on a three-dimensional reconstruction technology, and belongs to the technical field of computer vision and digital image forensics (multimedia information security). According to the method, an uncertainty perception geometric fusion module (UGFM) and a cross-scale query module (CSQM) are introduced, and a confidence map reflecting the reliability of geometric features is explicitly generated by calculating a second derivative (Laplacian operator) of a depth map, modulus length deviation of a normal vector and cross-modal consistency; generating a spatial gating weight to weight the features according to the spatial gating weight; during cross-scale interaction, cosine similarity is adopted to replace dot product to calculate an attention map, only the direction consistency of feature vectors is concerned, the influence of the die length is ignored, and higher robustness is achieved during multi-scale feature fusion. According to the method, the generalization ability in a cross-domain scene is remarkably improved, and higher convergence efficiency and boundary positioning precision are achieved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and digital image forensics (multimedia information security) technology, specifically involving an image tampering location method and system based on three-dimensional reconstruction technology. Background Technology

[0002] Image tampering localization technology aims to identify forged or modified areas in an image. With the development of deep learning technology, image editing software and generative models (such as diffusion-based generative algorithms) can generate visually extremely realistic forged images, making them difficult to distinguish with the naked eye. Therefore, automatic detection algorithms based on computer vision have become a research hotspot in this field.

[0003] Currently available deep learning-based two-dimensional feature extraction and anomaly detection solutions typically treat images as two-dimensional pixel matrices, utilize neural network models to extract spatial features, and locate tampered regions by analyzing discontinuities or abnormal patterns in the feature maps. Specific implementation methods mainly include the following categories: Local anomaly detection schemes based on convolutional neural networks (CNNs): These schemes primarily utilize the powerful local feature extraction capabilities of CNNs. The core idea is to extract RGB features or noise residual features from an image and search for statistically regular anomalies at the pixel level. For example, by comparing inconsistencies in noise distribution between the target region and the background region, or by detecting abrupt pixel changes at edges, it can be determined whether the region has undergone splicing or erasure processing.

[0004] Transformer-based long-range semantic consistency detection schemes: To compensate for the limited receptive field of CNNs, some existing techniques have introduced the Transformer architecture. These schemes utilize a self-attention mechanism to model the dependencies between long-distance pixels in an image. Their working principle is to compare the semantic relationships between different objects or regions in an image to discover semantic logical contradictions, thereby assisting in locating tampered regions.

[0005] Hybrid architecture based on multi-scale feature fusion: To simultaneously capture local details and global semantics, current state-of-the-art solutions typically employ a hybrid architecture. These solutions construct multi-branch encoders that extract shallow detail features and deep semantic features from the image separately, and then fuse two-dimensional features of different resolutions through feature pyramids or cascaded structures to capture tampering traces at different scales.

[0006] The aforementioned mainstream image tampering localization techniques (based on 2D features) have limited detection capabilities when faced with visually realistic, highly tampered images that conform to two-dimensional statistical laws, especially exhibiting poor generalization ability in cross-domain scenarios, easily leading to severe false negatives. Existing methods attempting to incorporate physical geometric cues either have narrow applicability due to reliance on specific assumptions, or, when attempting to utilize 3D reconstruction information, often result in unstable model training or even introduce interference leading to performance degradation, failing to truly and effectively utilize 3D geometric features.

[0007] Most existing deep learning models extract texture, noise residuals, or semantic context features only from the perspective of a two-dimensional image plane, ignoring the physical fact that an image is essentially a two-dimensional projection of a three-dimensional scene. With advancements in image editing and generation techniques, generated fake images have become highly similar to real images in terms of two-dimensional texture continuity and statistical features. However, they often contain logical errors that violate physical laws in terms of three-dimensional spatial structure (such as the continuity of depth levels, the perspective consistency between object surface normals and lighting). Current technologies lack the ability to perceive these three-dimensional geometric features and therefore fail to capture these implicit geometric inconsistencies, leading to detection failures.

[0008] Since tampered images typically lack true 3D geometric annotations, monocular 3D reconstruction algorithms must be used to estimate geometric information from RGB images. However, existing reconstruction algorithms often produce depth maps and normal maps containing inherent noise, artifacts, or blur (uncertainty) when dealing with complex tampered images. Current technologies lack effective mechanisms for evaluating and filtering the "reliability" of these geometric features. If these noisy, "imperfect" 3D features are forcibly fused with RGB features, the noise becomes an interfering signal, compromising the purity of the feature representation. This makes it impossible for the model to distinguish between tampering traces and reconstruction errors, thus affecting the final localization accuracy.

[0009] To address the aforementioned problems, this invention proposes an image tampering localization method and system based on three-dimensional reconstruction technology. Summary of the Invention

[0010] The purpose of this invention is to provide an image tampering localization method and system based on 3D reconstruction technology to solve the problem that most existing image tampering localization technologies only focus on the two-dimensional visual features of the image (such as pixel noise residuals and semantic inconsistencies), which leads to the inability to effectively identify complex tampered images that are visually realistic but have geometric logical conflicts in three-dimensional physical space (such as perspective errors, depth discontinuities, and abnormal normal vectors). At the same time, it solves the technical problem of how to overcome the noise and uncertainty brought by the reconstruction model itself when using monocular 3D reconstruction information as auxiliary clues, and achieve adaptive and robust fusion of 3D geometric features and 2D RGB features, thereby improving the localization accuracy of the model in cross-domain scenarios.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: A method for locating image tampering based on 3D reconstruction technology includes the following steps: S1. 3D Geometric Scene Reconstruction and Resampling: Input the RGB image to be detected, use a pre-trained monocular 3D reconstruction model to generate the corresponding 3D point cloud scene, and resample the 3D point cloud into a depth map spatially aligned with the original RGB image (denoted as ). D ) and surface normal vector map (denoted as N p ); among which, depth map ( D The values ​​in the surface normal map represent the absolute or relative depth of the pixel; N p The values ​​in the vector represent the surface orientation vector of the pixel in three-dimensional space; S2, Multimodal Feature Encoding: Construct an encoder with three parallel branches to extract RGB images and depth maps respectively. D ) and surface normal vector map ( N p The multi-scale features of ) are obtained for the i-th scale level; the corresponding RGB features are obtained respectively. ), deep features ( ) and normal vector features ( ); S3, Uncertainty-Aware Geometric Fusion (UGFM): Based on Depth Map ( D ) and surface normal vector map ( N p ), calculate the geometric confidence index, and generate spatial gating weights based on the geometric confidence index ( G i ), dynamically incorporating deep features ( ) and normal vector features ( ) merged into RGB features ( In ), the fusion features are obtained ( Z i ); S4. Cross-Scale Query and Decoding (CSQM): Employs a top-down decoding structure, using the features of the current decoding level as the query at each decoding level. Q The features of the next higher level serve as keys. K ) and value ( V The attention map is calculated based on cosine similarity and feature enhancement is performed. Finally, the tamper localization prediction mask is output. S5. Model Training and Output: The model is trained using a boundary-aware structural loss function to obtain a trained image tampering localization model. The trained model is then used to assign higher loss weights to pixels in the mask edge region, forcing the model to learn clear tampering boundaries and outputting the final tampering localization result.

[0012] Preferably, the monocular 3D reconstruction model used in S1 is MoGe2.

[0013] Preferably, the calculation of the geometric confidence index in S3 specifically includes: Calculating the depth confidence map: The second derivative of the depth map is calculated using the Laplacian operator to assess the drasticness (smoothness) of depth changes. The specific calculation formula is as follows:

[0014] in, C d This represents a depth confidence plot. Represents a depth map; D Represents a depth map; Calculate the confidence map of the normal vector: Based on the physical constraints of the unit vector, calculate the surface normal vector map ( N p The deviation between the modulus and the unit length 1 is calculated using the following formula:

[0015] in, C n Represents the confidence plot of the normal vector; N p Represents the surface normal vector diagram; Calculate the cross-modal consistency map: verify the consistency between depth and normal, using the depth map ( D ) Calculate and derive the normal vector ( N d ), and calculate its relationship with the surface normal vector diagram ( N p The cosine similarity of () is calculated using the following formula:

[0016] in, A dn Represents the cross-modal consistency graph; N d · denotes the derivation of the normal vector; · denotes the dot product operation.

[0017] Preferably, the calculation formula for the fusion feature in S3 is as follows:

[0018] in,Z i Indicates fusion characteristics; Represents RGB characteristics; Represents depth features; Represents the characteristics of the normal vector; G i This represents the spatial gating weight, and its value ranges from 0 to 1; w d and w n Represents learnable scalar weights; This indicates element-wise multiplication.

[0019] Preferably, the cosine similarity in S4 is used to eliminate the influence of feature amplitude, allowing the model to focus on the directional consistency of features to capture semantic anomalies in the tampered region; the formula for calculating the cosine similarity is:

[0020] Wherein, the superscript T denotes transposed convolution; This represents the similarity distance scheduling parameter.

[0021] Preferably, S4 further includes a gated feedforward step, which uses a network structure containing depthwise separable convolutions to refine the attention-enhanced features.

[0022] An image tampering localization system based on 3D reconstruction technology, the system comprising: 3D geometric scene reconstruction and resampling module: This module takes an RGB image to be detected as input, generates a corresponding 3D point cloud scene using a pre-trained monocular 3D reconstruction model, and resamples the 3D point cloud into a depth map spatially aligned with the original RGB image (denoted as ). D ) and surface normal vector map (denoted as N p ); among which, depth map ( D The values ​​in the surface normal map represent the absolute or relative depth of the pixel; N p The values ​​in the vector represent the surface orientation vector of the pixel in three-dimensional space; Multimodal feature encoding module: used to construct an encoder with three parallel branches to extract RGB images, depth maps, etc. D ) and surface normal vector map ( N p The multi-scale features of ) are obtained for the i-th scale level; the corresponding RGB features are obtained respectively. ), deep features ( ) and normal vector features ( ); Uncertainty-Aware Geometric Fusion (UGFM) module: for use with depth maps ( D ) and surface normal vector map ( N p ), calculate the geometric confidence index, and generate spatial gating weights based on the geometric confidence index ( G i ), dynamically incorporating deep features ( ) and normal vector features ( ) merged into RGB features ( In ), the fusion features are obtained ( Z i ); Cross-Scale Query and Decode (CSQM) module: Employs a top-down decoding structure, utilizing the features of the current decoding level as the query at each decoding level. Q The features of the next higher level serve as keys. K ) and value ( V The attention map is calculated based on cosine similarity and feature enhancement is performed. Finally, the tamper localization prediction mask is output. Model training and output module: The model is trained using a boundary-aware structural loss function to obtain a trained image tampering localization model; the trained model is then used to assign higher loss weights to pixels in the mask edge region, forcing the model to learn clear tampering boundaries, and outputting the final tampering localization result.

[0023] The present invention further protects a computer device, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the instruction, program, code set or instruction set being loaded and executed by the processor to implement the above-mentioned image tampering location method based on three-dimensional reconstruction technology.

[0024] The present invention further provides a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the instruction, program, code set, or instruction set is loaded and executed by a processor to implement the above-described image tampering location method based on three-dimensional reconstruction technology.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention significantly improves the generalization ability and anti-interference robustness in cross-domain scenarios. Compared with the prior art, this invention significantly improves the F1 score on unseen test data (OOD) (e.g., an improvement of about 14%), and can still maintain high accuracy when the image is subjected to noise, blur or social media compression. This invention adopts a method that combines auxiliary branch based on monocular 3D reconstruction with uncertainty-aware geometric fusion, introduces more physically universal 3D geometric features (depth and normal), automatically identifies and suppresses geometric reconstruction noise caused by image quality degradation, and ensures that the model only uses reliable physical conflict clues for reasoning, thereby maintaining stability in complex and variable scenarios.

[0026] (2) This invention has higher convergence efficiency and boundary localization accuracy. Compared with the prior art, this invention can achieve the best results with fewer training rounds (e.g., 20 rounds compared to 50 rounds in the prior art), and can generate localization masks with sharper edges that fit the real objects. This invention adopts a combination of cross-scale query module and dynamic boundary structure loss. Through the cosine similarity attention mechanism, global semantic information is efficiently injected into high-resolution features, forcing the model to gradually focus on high-difficulty boundary pixels during training, thereby accelerating the convergence of the model and optimizing the detail performance. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings involved in the embodiments are now briefly described. Obviously, the drawings in the following description are merely illustrative of some embodiments of the present invention. For those skilled in the art, other forms of drawings can be constructed based on these drawings without creative effort.

[0028] Figure 1 This is an overall architecture diagram of the image tampering localization model based on three-dimensional reconstruction technology proposed in Embodiment 1 of the present invention; Figure 2 This is the visualization of gated reliability mentioned in Embodiment 1 of the present invention; Figure 3 This is a comparison image of the final model visual effects mentioned in Embodiment 1 of the present invention. Detailed Implementation

[0029] 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.

[0030] This invention proposes an image tampering localization method and system based on three-dimensional reconstruction technology, the core objective of which is: 1. Overcoming the limitations of 2D feature detection: By introducing monocular 3D reconstruction technology, the depth map and surface normal vector of the scene are recovered from the input RGB image. The physical consistency of the 3D geometric space is used as an auxiliary supervision signal, which can effectively identify complex tampered areas that are continuous in 2D texture but have discontinuities or anomalies in 3D structure.

[0031] 2. Addressing the interference of noise in 3D reconstruction: A geometric fusion mechanism with uncertainty perception capability is designed to dynamically evaluate the reliability (confidence) of geometric features in different regions based solely on the reconstruction results, without manual annotation. This mechanism adaptively suppresses low-quality, high-noise geometric features and enhances the weight of high-confidence geometric cues, thereby achieving robust fusion of 3D geometric information and 2D RGB information and preventing reconstruction errors from misleading model judgments.

[0032] 3. Enhance cross-domain generalization ability and boundary accuracy: By designing a cross-scale query mechanism, effective interaction between high-resolution detail features and low-resolution semantic features is promoted, ensuring the effective propagation of geometric consistency information across different scales. Combined with a specific structured loss function, the generalization performance of the model under different data distributions is ultimately improved, and the localization accuracy of the edge contours of tampered regions is significantly enhanced.

[0033] The following description, in conjunction with the accompanying drawings and specific examples, illustrates the image tampering localization method and system based on three-dimensional reconstruction technology proposed in this invention.

[0034] Example 1: This example proposes an image tampering localization method (GFrame) based on 3D reconstruction technology. The core of this method lies in introducing an uncertainty-aware geometric fusion module (UGFM) and a cross-scale query module (CSQM) to quantify the reliability of geometric features and perform dynamic feature fusion accordingly. The overall architecture is as follows: Figure 1 As shown.

[0035] The specific implementation steps of the proposed method are as follows: Step 1: 3D geometric scene reconstruction and resampling: The input is an RGB image to be detected, and a pre-trained monocular 3D reconstruction model (MoGe2) is used to generate a 3D point cloud scene. Then, the point cloud is resampled into a depth map (denoted as ) aligned with the original RGB image space. D ) and surface normal vector map (denoted as N p ).in, D The values ​​in the text represent the absolute or relative depth of a pixel. N pThe numerical value in the vector represents the surface orientation vector of a pixel in three-dimensional space.

[0036] Step 2: Multimodal Feature Encoding Construct an encoder with three parallel branches to extract RGB images and depth maps respectively. D and normal vector graph N p Multi-scale features. For the first i Each scale level yields RGB features. Depth features and normal vector features .

[0037] Step 3: Uncertainty-Aware Geometric Fusion (UGFM) This step is the core of the invention. Due to the noise in the reconstruction process, the invention dynamically filters features by calculating three geometric confidence indices.

[0038] (1) Calculate the geometric confidence index Define three confidence plots: depth confidence Normal vector confidence and cross-modal consistency The specific calculation formula is as follows: Depth confidence ( ): Using the Laplace operator ( Calculate the second derivative of the depth map to assess the drasticness (smoothness) of depth variations:

[0039] This formula indicates that the more drastic the depth changes (such as noise or unnatural faults), the more likely the problem will occur. The smaller the value, the less reliable it is.

[0040] Normal vector confidence ( (): Physical constraints based on unit vectors. Calculate the predicted normal vector. The deviation of the modulus from the unit length 1:

[0041] This formula shows that the greater the deviation of the modulus from 1, the worse the reconstruction quality. C n The smaller the value.

[0042] Cross-modal consistency ( ): Verify the consistency between depth and normal. This is done using the depth map. D Calculate and derive the normal vector N d And calculate its relationship with the predicted normal vector. N pCosine similarity:

[0043] The formula normalizes the result to the range [0, 1]. The closer the value is to 1, the more geometrically consistent the depth and normal are, and the more reliable the information is.

[0044] (2) Dynamic gating and feature fusion Based on the above confidence level, calculate the spatial gating weight. G i The geometric features are then fused into the RGB features. The fusion formula is as follows:

[0045] Among them, Z i These are the characteristics after fusion; G i These are the gate weights (ranging from 0 to 1) generated by a convolutional network from the concatenated features and confidence map. w d and w n These are learnable scalar weights; This indicates element-wise multiplication.

[0046] Using this formula, the model can automatically... G i Smaller regions (unreliable regions) suppress the injection of geometric features, avoiding noise interference, and the visual effect is as follows: Figure 2 As shown.

[0047] Step 4: Cross-Scale Query and Decoding (CSQM) To recover high-resolution details, a top-down decoding structure is employed. At each level, feature enhancement is performed using a cross-scale query module (CSQM). Input: High-resolution features of the current level as query (Q), and low-resolution features of the previous level as key-value pairs (K / V).

[0048] Attention Calculation: Cosine similarity is used to calculate the attention map, instead of the traditional dot product. The formula is briefly described as follows:

[0049] Using cosine similarity can eliminate the influence of feature magnitude, allowing the model to focus on the directional consistency of features, thereby more accurately capturing semantic anomalies in tampered regions.

[0050] Gated feedforward: The fused features are further refined through a gated network containing depthwise separable convolutions.

[0051] Step 5: Loss Function and Model Training The model outputs the final tampering localization mask. During training, a boundary-aware structural loss method is used, assigning higher loss weights to pixels in the mask edge regions, forcing the model to learn clear tampering boundaries, resulting in an overall visual effect as shown. Figure 3 As shown.

[0052] Example 2: Based on Embodiment 1, but with a difference, the image tampering location method based on three-dimensional reconstruction technology proposed in this invention is compared with commonly used image tampering location methods in the field, and the specific results are shown in Table 1.

[0053] Table 1. Comparison of in-domain and out-of-domain effects of common methods

[0054] As shown in Table 1, our solution outperforms existing baselines in both overall out-of-domain and in-domain performance.

[0055] Existing multimodal image processing or tampering detection methods typically employ simple channel stitching or element-wise addition with fixed weights when fusing features from different modalities (such as RGB with noise flow or RGB with depth). This approach assumes that all input modalities are equally reliable across all regions, ignoring the fact that 3D reconstruction information in forged images often contains local noise and reconstruction errors. This application proposes an Uncertainty-Aware Geometric Fusion Mechanism (UGFM), introducing a dynamic gating mechanism based on physical geometric constraints. Unlike existing direct fusion techniques, this invention first explicitly generates a "confidence map" reflecting the reliability of geometric features by calculating the second derivative of the depth map (Laplacian operator), the magnitude deviation of the normal vector, and cross-modal consistency. Based on this, spatial gating weights are generated to weight the features. The core technical difference lies in the fact that this invention adds a "quality assessment and screening" step based on physical laws, technically solving the problem of low-quality reconstructed features interfering with the main network.

[0056] Furthermore, existing Transformer-based tamper detection models generally employ standard scaled dot products when calculating attention, and often use simple skip connections similar to U-Net or direct upsampling concatenation during the decoding stage. This application proposes a cosine similarity-based cross-scale query decoding architecture (CSQM), which uses cosine similarity to replace dot products in calculating the attention map during cross-scale interactions and designs a specific query structure for "high-resolution features querying low-resolution context." Existing dot product attention techniques are sensitive to feature magnitude, easily leading to training instability; while the technique used in this invention (cosine similarity) only focuses on the directional consistency of feature vectors, ignoring the influence of magnitude, and has stronger robustness in multi-scale feature fusion. Simultaneously, the active query mechanism of this invention can more effectively inject global semantics into local details than existing passive concatenation.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for locating image tampering based on three-dimensional reconstruction technology, characterized in that, Includes the following steps: S1. 3D Geometric Scene Reconstruction and Resampling: Input the RGB image to be detected, use the pre-trained monocular 3D reconstruction model to generate the corresponding 3D point cloud scene, and resample the 3D point cloud into a depth map and a surface normal vector map spatially aligned with the original RGB image; where the values ​​in the depth map represent the absolute or relative depth of the pixel; the values ​​in the surface normal vector map represent the surface orientation vector of the pixel in 3D space. S2. Multimodal feature encoding: Construct an encoder with three parallel branches to extract multi-scale features from the RGB image, depth map, and surface normal vector map, respectively; for the i-th scale level, obtain the corresponding RGB features, depth features, and normal vector features, respectively; S3. Uncertainty-aware geometric fusion: Based on the depth map and surface normal vector map, calculate the geometric confidence index, generate spatial gating weights according to the geometric confidence index, and dynamically fuse the depth features and normal vector features into the RGB features to obtain the fused features; S4, Cross-scale query and decoding: A top-down decoding structure is adopted. At each decoding level, the features of the current level are used as the query and the features of the previous level are used as the key and value. An attention map is calculated based on cosine similarity and feature enhancement is performed. Finally, the tamper location prediction mask is output. S5. Model Training and Output: The model is trained using a boundary-aware structural loss function to obtain a trained image tampering localization model. The trained model is then used to assign higher loss weights to pixels in the mask edge region, forcing the model to learn clear tampering boundaries and outputting the final tampering localization result.

2. The method according to claim 1, characterized in that, The monocular 3D reconstruction model used in S1 is MoGe2.

3. The method according to claim 1, characterized in that, The calculation of the geometric confidence index described in S3 specifically includes: Calculating the depth confidence map: The second derivative of the depth map is calculated using the Laplace operator to assess the drasticness of depth changes. The specific calculation formula is as follows: in, C d This represents a depth confidence plot. Represents a depth map; D Represents a depth map; Calculating the confidence map of the normal vector: Based on the physical constraint of the unit vector, the deviation of the magnitude of the surface normal vector map from the unit length 1 is calculated. The specific calculation formula is as follows: in, C n Represents the confidence plot of the normal vector; N p Represents the surface normal vector diagram; Calculate the cross-modal consistency map: verify the consistency between depth and normal vectors. Derive the normal vectors from the depth map and calculate their cosine similarity to the surface normal vector map. The specific calculation formula is as follows: in, A dn Represents the cross-modal consistency graph; N d · denotes the derivation of the normal vector; · denotes the dot product operation.

4. The method according to claim 3, characterized in that, The calculation formula for the fusion feature described in S3 is as follows: in, Z i Indicates fusion characteristics; Represents RGB characteristics; Represents depth features; Represents the characteristics of the normal vector; G i This represents the spatial gating weight, and its value ranges from 0 to 1; w d and w n Represents learnable scalar weights; This indicates element-wise multiplication.

5. The method according to claim 1, characterized in that, The cosine similarity described in S4 is used to eliminate the influence of feature amplitude, allowing the model to focus on the directional consistency of features in order to capture semantic anomalies in tampered regions; the formula for calculating the cosine similarity is: Wherein, the superscript T denotes transposed convolution; This represents the similarity distance scheduling parameter.

6. The method according to claim 1, characterized in that, The S4 also includes a gated feedforward step, which uses a network structure containing depthwise separable convolutions to refine the attention-enhanced features.

7. An image tampering localization system based on three-dimensional reconstruction technology for implementing the method of any one of claims 1-6, characterized in that, The system includes: 3D geometric scene reconstruction and resampling module: It is used to take the RGB image to be detected as input, generate the corresponding 3D point cloud scene using a pre-trained monocular 3D reconstruction model, and resample the 3D point cloud into a depth map and a surface normal vector map that are spatially aligned with the original RGB image; where the values ​​in the depth map represent the absolute or relative depth of the pixel; the values ​​in the surface normal vector map represent the surface orientation vector of the pixel in 3D space. Multimodal feature encoding module: used to construct an encoder with three parallel branches to extract multi-scale features from RGB image, depth map and surface normal vector map respectively; for the i-th scale level, the corresponding RGB features, depth features and normal vector features are obtained respectively; Uncertainty-aware geometric fusion module: It is used to calculate the geometric confidence index based on the depth map and surface normal vector map, generate spatial gating weights based on the geometric confidence index, and dynamically fuse the depth features and normal vector features into the RGB features to obtain the fused features; Cross-scale query and decoding module: It adopts a top-down decoding structure. At each decoding level, it uses the features of the current level as the query and the features of the previous level as the key and value. It calculates the attention map based on cosine similarity and performs feature enhancement. Finally, it outputs the tamper location prediction mask. Model training and output module: The model is trained using a boundary-aware structural loss function to obtain a trained image tampering localization model; the trained model is then used to assign higher loss weights to pixels in the mask edge region, forcing the model to learn clear tampering boundaries, and outputting the final tampering localization result.

8. A computer device, characterized in that, The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set, or instruction set, and the instruction, program, code set, or instruction set is loaded and executed by the processor to implement the image tampering location method based on three-dimensional reconstruction technology as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, at least one program, code set, or instruction set, which is loaded and executed by a processor to implement the image tampering location method based on three-dimensional reconstruction technology as described in any one of claims 1-6.

Images