Image tampering positioning method and system based on label decoupling

Abstract

The invention discloses an image tampering positioning method and system based on label decoupling, and belongs to the technical field of image evidence obtaining and deep learning. According to the method, a tampering mask is divided into a body area (an easy-to-predict interior) and a detail area (a difficult-to-predict boundary) through a label decoupling strategy, a double-span hourglass Transform encoder (DH-ViT) and a bidirectional decoupling pyramid decoder (BDPD) are designed to realize separation, extraction and optimization of two types of features, and intelligent fusion of the features is realized through a body-detail adaptive relation module (BARM). The limitation of the prior art is comprehensively broken through from three levels of label design, network architecture and feature fusion, and the problems that the positioning precision and the boundary fine positioning capability are difficult to guarantee at the same time and the uncertainty estimation of a tampered area is insufficient in the prior art when facing an image in a complex scene are effectively solved. And the problems of boundary blur, false detection or missing detection of a detection result are solved.

Description

Technical Field

[0001] This invention belongs to the field of image forensics and deep learning technology, specifically relating to an image tampering location method and system based on label decoupling. Background Technology

[0002] Image Manipulation Localization (IML) is a key technology in digital image forensics, aiming to identify manipulated areas in images at the pixel level to ensure the authenticity of multimedia content. With the widespread use of deepfakes and image editing tools, IML is becoming increasingly important in areas such as news verification, forensic evidence collection, and social media content moderation.

[0003] The core challenges faced by traditional IML methods include:

[0004] Although existing methods have attempted to introduce edge information as auxiliary supervision into the model, they have not fundamentally solved the problems of boundary imbalance and complex boundary modeling. Therefore, a new technical approach is urgently needed to break through this bottleneck.

[0005] To address these adjustments, existing technologies have explored various approaches, with representative solutions including: Addressing the issue of ambiguous boundaries within the field: The paper "Imagemanipulation detection by multi-view multi-scale supervision." presented in the Proceedings of the IEEE / CVF International Conference on Computer Vision (ICCV), 2021," proposes a multi-view, multi-scale learning scheme (MVSS-Net) based on edge supervision. This scheme posits that edge features play a crucial guiding role in tamper detection and proposes a multi-view, multi-scale supervision mechanism to fuse edge information with noise distribution features to improve boundary detection accuracy. Its core idea is to extract features from multiple views (such as noise distribution views and edge views) and supervise the model's training at multiple scales to capture the global structure and local edge details of the tampered region.

[0006] The paper "TBNet: A two-stream boundary-aware network for generic image manipulation localization" by Zan Gao, Chao Sun, Zhiyong Cheng, Weili Guan, Anan Liu, and Meng Wang in IEEE Transactions on Knowledge and Data Engineering, 35(7):7541–7556, 2022" proposes a boundary-aware two-stream network scheme (TBNet). This scheme addresses the challenge of boundary localization by proposing a two-stream network architecture. A dedicated boundary-aware branch is designed to focus on extracting artifact features of tampered boundaries, and a boundary-aware loss function is used to enhance the supervision of key boundary pixels, thereby achieving general image manipulation localization.

[0007] The paper "Edge Distribution Guidance and Contrastive Learning (EC-Net)" by Qixian Hao, Ruyong Ren, Kai Wang, Shaozhang Niu, Jiwei Zhang, and Maosen Wang, proposes an edge distribution guidance and contrastive learning scheme. EC-Net generates prior edge information through an independent edge prediction module, guides the backbone network to focus on boundary regions using edge distribution, and combines contrastive learning to improve the discriminative ability of features, thereby improving the localization effect of tampered boundaries.

[0008] Addressing the inherent conflict issues of out-of-domain tags: The paper "Label decoupling framework for salient object detection." by Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian in IEEE / CVFConference on Computer Vision and Pattern Recognition (CVPR), 2020" proposes a label decoupling framework (LDF). This scheme decouples the binary labels of salient objects into core region labels and edge region labels, and achieves the separation learning of core structure and edge details through differential supervision. However, it uses a unidirectional pyramid encoder, which leads to mutual interference between core features and edge features during feature extraction. Furthermore, the boundary features of the SOD task are relatively simple and cannot be directly transferred to the complex artificially hidden boundary scenes in the IML task.

[0009] The paper "Rui Zhao, Yuetong Li, Qing Zhang, and Xinyi Zhao. 'Bilateral decoupling complementarity learning network for camouflaged object detection. Knowledge-Based Systems, 314:113158, 2025'" proposes a bilateral decoupling complementary learning network (BDCL-Net). This scheme divides the labels of camouflaged targets into foreground cores and edge transition regions through bilateral decoupling and optimizes feature extraction using a complementary learning strategy. However, it adopts a hard segmentation strategy, completely losing high-level semantic guidance, resulting in limited generalization ability in complex boundary scenes and an inability to adapt to the diversity and concealment of tampered boundaries in IML tasks.

[0010] The above-mentioned schemes pay extra attention to boundary information in model supervision or structural design, but they still have certain limitations and shortcomings. In order to solve the above problems, this invention proposes an image tampering localization method and system based on label decoupling. Summary of the Invention

[0011] The purpose of this invention is to provide a label-decoupled image tampering localization method and system to solve the following problems existing in current image tampering localization technologies: (1) Within the tampered area, the boundary imbalance problem at the label level has not been fundamentally solved. It still relies on binary labels or simple hard segmentation labels, which cannot reflect the difficulty difference of pixels within the boundary area. (2) Outside the tampered area, there is an inherent conflict in the feature extraction architecture. The extraction of global structural features and edge detail features interfere with each other and are difficult to optimize at the same time.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: A label-decoupled image tampering localization method includes the following steps: The tampered image to be detected is acquired, and the original binary tampering mask corresponding to the tampered image is decomposed into a body map and a detail map through a label decoupling (LD) process. The body map is used to represent the main structure of the tampered region, and the detail map is used to represent the boundary details of the tampered region. The processed tampered image is input into the Dual-Span Hourglass Vision Transformer (DH-ViT) encoder for feature extraction, obtaining body feature streams and detail feature streams adapted to the body image and detail image respectively; The body feature stream and detail feature stream are respectively input into two independent paths in the Bi-directional Decoupling Pyramid Decoder (BDPD) for parallel refinement and optimization, resulting in optimized body features and detail features; The optimized body features and detail features are input into the Body-Detail Adaptive Relation Module (BARM) for adaptive fusion to generate a final prediction mask that combines structural integrity and boundary accuracy. A multi-task loss function is used to jointly supervise the body prediction corresponding to body features, the detail prediction corresponding to detail features, and the final prediction mask to ensure that the components of the model are optimized collaboratively.

[0013] Preferably, the tag decoupling process specifically includes: Decomposition is achieved by performing morphological operations on the original binary tampering mask, specifically by obtaining the body image through dilation and obtaining the detail image by performing a difference operation between the original binary tampering mask and the body image.

[0014] Preferably, the dual-span hourglass visual Transformer encoder adopts an hourglass-shaped structure, including a sequentially connected downsampling path and an upsampling path; The downsampling path is used to extract coarse-grained, high-semantic body feature streams; the upsampling path is used to extract fine-grained, high-resolution detail feature streams. The fusion and transfer of features in the downsampling path and the upsampling path are achieved through a double-span skip connection.

[0015] Preferably, the bidirectional decoupled pyramid decoder includes a body grouping fusion module and a detail grouping fusion module arranged in parallel; The body grouping and fusion module is used to refine the body feature stream at multiple scales, capture the global structural information of the tampered region through multi-branch convolution operations, and output the optimized body features. The detail grouping and fusion module is used to refine the detail feature stream at multiple scales. It captures subtle traces of boundary tampering through multi-branch convolution operations and outputs optimized detail features.

[0016] Preferably, the adaptive fusion operation of the body-detail adaptive relationship module specifically includes the following: Local attention modeling is performed on the optimized body features and detail features of the input to highlight key regions related to tampering; The attention-modeled features are converted into query vectors, key vectors, and value vectors, and the association between the optimized body features and detail features is calculated through the attention mechanism. Graph convolutional networks are used to perform graph convolution operations on the associated features to model the global relationships between features. By fusing the features processed by graph convolution through multiplication and addition operations, a fused feature is obtained to generate the final tamper location prediction mask.

[0017] Preferably, the multi-task loss function includes a supervised loss for predicting the body branch, a supervised loss for predicting the detail branch, and a supervised loss for predicting the final prediction mask.

[0018] A label-decoupled image tampering localization system includes: The image acquisition module is used to acquire the tampered image to be detected; The label decoupling module is used to decompose the original binary tampering mask corresponding to the tampered image into a body image and a detail image through the label decoupling process; The feature encoding module includes a dual-span hourglass visual Transformer encoder, which is used to extract features from tampered images and output body feature streams and detail feature streams. The feature decoding optimization module includes a bidirectional decoupled pyramid decoder, which is used to refine and optimize the body feature stream and detail feature stream in parallel through two independent paths to obtain optimized body features and detail features; The feature fusion module includes a body-detail adaptive relationship module, which is used to adaptively fuse optimized body features and detail features to generate the final tamper localization prediction mask; The training supervision module is used to jointly supervise the model based on the body prediction corresponding to the body features, the detail prediction corresponding to the detail features, and the final prediction mask, and then construct a multi-task loss function to jointly supervise the training of the model.

[0019] Preferably, the bidirectional decoupled pyramid decoder specifically includes: The body grouping and fusion submodule is used to refine the body feature flow in a multi-scale structure. The detail grouping and fusion submodule is used to refine the detail feature stream at multiple scales. The body grouping fusion submodule and the detail grouping fusion submodule are structurally independent and work in parallel.

[0020] 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 tag decoupling.

[0021] 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 tag decoupling.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: Compared to existing technologies, the image tampering localization method based on label decoupling proposed in this invention has higher boundary localization accuracy, more efficient feature learning, better fusion effect, and stronger robustness. Each advantage is supported by targeted innovative technical means, specifically including: The improved accuracy of boundary localization stems from the adoption of a label decoupling (LD) process based on Euclidean distance transform (EDT). Unlike existing binary labels or hard segmentation labels, this process flexibly decomposes the original mask into a body map that focuses on the internal structure and a detail map that focuses on the boundary details, providing differentiated supervision for the model. This allows the difficult-to-predict boundary regions to obtain sufficient learning weights and avoids them being masked by a large number of easily predictable internal pixels. The high efficiency of feature learning is due to the dual-span hourglass vision Transformer (DH-ViT) encoder, whose symmetrical downsampling-upsampling structure and dual-span skip connections completely separate the extraction process of body features (global structure) and detailed features (high-frequency boundary information), solving the feature extraction conflict problem existing in pyramid encoders or dual-stream architectures in the prior art. The optimization of the fusion effect is attributed to the Body-Detail Adaptive Relationship Module (BARM), which uses a dynamic fusion mechanism of "local attention + graph convolution (GCN)" instead of static splicing or weighted summation in existing technologies. This allows for adaptive modeling of the relationship between the two types of features, taking into account both the integrity of the regional structure and fine-grained boundary localization. Multi-task joint supervision (body, details, and final prediction are all independently supervised) and separate feature optimization further make the model more resistant to disturbances such as noise, blur, compression, and image degradation after propagation on social platforms. Its robustness is significantly better than existing technologies that rely on single supervision or static fusion. Attached Figure Description

[0023] 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.

[0024] Figure 1 This is a diagram illustrating the overall structure of the image tampering localization method and system based on tag decoupling proposed in this invention. Detailed Implementation

[0025] 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.

[0026] This invention proposes a label-decoupled image tampering localization method and system. Its core objective is to specifically address the key shortcomings of existing image tampering localization (IML) technologies, achieving more accurate, robust, and practically applicable tampering boundary localization and region detection. Specifically, it includes: 1. Solving the problem of imbalanced boundary pixels at the label level: In response to the shortcomings of existing technologies that rely on fixed binary labels and cannot distinguish the difference in difficulty between boundary (difficult to predict) and interior (easy to predict) pixels within a tampered area, a label decoupling strategy is adopted to provide the model with more targeted supervision signals, allowing the model to focus on fine-grained feature learning in complex boundary areas, and avoiding the "drowning" of difficult-to-predict boundary pixels by a large number of easily predictable interior pixels.

[0027] 2. Eliminate inherent conflicts in the feature extraction process: To address the mutual interference problem in the extraction of global structural features (requiring depth downsampling) and boundary detail features (requiring the preservation of high-frequency information) by traditional pyramid encoders, a dedicated network architecture is designed to achieve the separate extraction and independent optimization of the two types of features, while simultaneously meeting the dual requirements of the integrity of the tampered region structure and the accuracy of boundary localization.

[0028] 3. Achieve adaptive and intelligent feature fusion: To address the shortcomings of existing technologies that use simple splicing or weighted summation for fusion and cannot dynamically model the complementary relationship between global structure and boundary details, an adaptive fusion mechanism is constructed to balance local fine-grained correlations and global structural consistency, so that the final prediction has both regional integrity and boundary accuracy.

[0029] 4. Enhance the robustness of the model in complex scenarios and real-world applications: Addressing the shortcomings of existing technologies, such as significant performance degradation under noise, blur, compression, and other disturbances, and difficulty in adapting to image quality degradation scenarios after dissemination on social media platforms, strengthen the model's resistance to various interferences to ensure that it can still stably perform positioning performance in real-world application scenarios.

[0030] The following description, in conjunction with the accompanying drawings and specific examples, illustrates the image tampering localization method and system based on label decoupling proposed in this invention.

[0031] Example 1: This example proposes a label-based decoupling method and system for image tampering localization, and more specifically, an hourglass label decoupling network (HLD-Net) to solve the boundary imbalance problem in image manipulation localization (IML) tasks, achieving the dual goals of structural integrity of the tampered region and accurate boundary localization.

[0032] The overall architecture of HLD-Net adopts an end-to-end design of "label decoupling - feature separation and extraction - independent optimization - adaptive fusion", as shown in Figure 1 (HLD-Net overall architecture diagram): The symbols on the left side of the diagram are fundamental to understanding the architecture, representing the core operations / components in the network: Conv 5×5 / 3×3 / 1×1: Convolutional layers of different sizes are used to extract features from different receptive fields; Addition (⊕): Element-level addition operation, used for feature fusion; Concatenation(C): Feature concatenation operation, which merges multiple sets of features according to the channel dimension; Conv+BN+ReLU: A combined module of "convolution + batch normalization + ReLU activation" used for feature transformation and activation; Upsampling (UP): An upsampling operation used to restore the resolution of the feature map; Color-coded branches: pink represents the body branch (processing tampered region structural features), green represents the detail branch (processing tampered boundary features), orange represents fusion features, and purple represents the supervision signal.

[0033] Its core process is as follows: Input a 512×512 image to be detected, and decompose the original binary tampering mask into a body map and a detail map through a label decoupling (LD) process to provide differential supervision signals; The dual-span hourglass vision transformer (DH-ViT) is used to extract features from the input image and output two types of feature streams that are adapted to the body image and the detail image, respectively. The Bi-directional Decoupling Pyramid Decoder (BDPD) performs parallel refinement and optimization of two types of feature streams through two independent paths; The Body-Detail Adaptive Relation Module (BARM) fuses optimized body features and detail features to generate a final prediction mask that combines structural integrity and boundary accuracy. A multi-task loss function is used to jointly supervise body prediction, detail prediction, and final prediction to ensure that the various components of the model are optimized collaboratively.

[0034] The above architecture fundamentally solves the core defects of traditional IML models, such as difficulty in predicting boundary pixels, mutual interference in feature extraction, and simple fusion methods, through decoupling design at the label level, separation and extraction mechanism at the network level, and adaptive modeling at the fusion level.

[0035] based on Figure 1 Its core modules and processes can be broken down as follows: Figure 1 The architecture flow shown is as follows: The process involves "input → encoder feature extraction → group fusion and refinement → adaptive fusion → prediction + supervision"; details of each module are as follows: 1. Input and Encoder: Two-Span Hourglass Vision Transformer Input: The original image in the upper left corner of the image (example: people on the grass) is the tampered image to be detected.

[0036] Double-span hourglass visual Transformer (large module at the top of the image): Structure: Shaped like an hourglass, it consists of multiple sets of Transformer blocks (pink, light pink, light blue, and cyan blocks in the image) connected in series, achieving the separation feature extraction of "downsampling abstract structure + upsampling to restore details".

[0037] Double-span jump connection: Feature transfer across blocks is achieved through ⊕ (addition) - downsampling path (pink branch) extracts body features (coarse-grained, high semantics, corresponding to the core structure of the tampered region); upsampling path (green branch) extracts detail features (fine-grained, high resolution, corresponding to the tampered boundary).

[0038] 2. Group fusion module (dashed box on the right side of the figure) Function: Refines the body / detail features output by the encoder at multiple scales, enhancing the specificity of the features.

[0039] Operation: Includes components such as Conv 3×3 / 5×5, Conv+BN+ReLU, and upsampling (UP). It fuses multi-scale features through "concatenation (C) + addition (⊕)" to ultimately output two types of optimized features: f_body: Optimization features of body branches (focusing on tampering with region structure); f_detail: Optimization features of detail branches (focusing on tampering with boundary details).

[0040] 3. BGF / DGF module: Grouping and fusion submodule BGF (Body Group Fusion, pink block): A module for refining body branches. It captures the global structure of the tampered region through multi-branch convolution and outputs more complete body features.

[0041] DGF (Detail Group Fusion, Cyan Block): A refinement module for detail branches. It captures subtle alterations to boundaries through multi-branch convolutions, outputting more accurate detail features.

[0042] The two processes work in parallel, handling body / detail features separately, thus avoiding mutual interference between the two types of features during optimization.

[0043] 4. Body-Detail Adaptive Relationship Module (BARM, dashed box below the image) Input: BGF / DGF output f_body (cyan block) and f_detail (pink block).

[0044] Operating procedures: Local Attention: Models attention for two types of features to highlight key regions relevant to tampering; Attention computation: features are converted into queries (Q), keys (K), and values ​​(V), and the relationship between the two types of features is captured through an attention mechanism; Graph Convolutional Networks (GCNs): Perform graph convolutions on attention features to model the global relationships between features; Fusion output: The final fused feature f_out is obtained by fusing the processed features through multiplication and addition operations.

[0045] 5. Prediction and Monitoring (bottom center of the diagram) Prediction: After processing, the fused feature f_out outputs the tampering localization result (the three result images with black backgrounds in the figure: the left is the detail branch prediction, the middle is the final prediction, and the right is the body branch prediction).

[0046] Supervision (purple arrow): The output of the body branch, detail branch, and final fusion are supervised and trained according to the "body graph, detail graph, and original truth mask" respectively, to ensure that each module is optimized in a coordinated manner.

[0047] This invention proposes a label decoupling (LD) image tampering localization process, a dual-span hourglass visual Transformer (DH-ViT) encoder, and a body-detail adaptive relation module (BARM), which together constitute a complete technology chain of "label differentiation - feature separation and extraction - adaptive fusion". Compared with existing technologies, the core difference lies in the innovative design of specific technical means, rather than simple functional optimization: At the label modeling level, existing technologies (such as MVSS-Net and TBNet) all use binary labels or simple edge supervision labels, only distinguishing between "tampered / untampered" or "edge / non-edge," without considering the difference in prediction difficulty of pixels within the tampered region. Cross-domain label decoupling schemes (such as LDF in the SOD domain) use hard segmentation strategies, losing semantic information and failing to adapt to the complex boundaries of IML. This invention is a non-binary label decoupling technology based on Euclidean distance transform (EDT). By normalizing the distance map, the original mask is decomposed into a body map (focusing on easily predictable internal regions) and a detail map (focusing on difficult-to-predict boundary regions). Technically, it achieves flexible label decomposition through "element-wise multiplication" rather than hard segmentation or binary labeling, providing differentiated supervision to the model from the source. This is a label modeling method not addressed in existing technologies.

[0048] At the feature extraction and fusion architecture level, existing encoders (such as the pyramid encoder of PSCC-Net and the two-stream architecture of TBNet) either suffer from an inherent conflict between "downsampling abstraction and upsampling detail preservation" or the two-path features still share some extraction modules, making complete separation impossible. Fusion methods (such as feature concatenation in EC-Net and weighted summation in MVSS-Net) are all static and simple fusions, failing to model the dynamic relationships between features. This invention proposes the hourglass-shaped dual-span architecture of DH-ViT, which, through a 6-stage symmetrical design and dual-span skip connections, allows the downsampling path to specifically extract body features and the upsampling path to specifically recover detailed features, separating the extraction processes of the two types of features from the architectural source. Simultaneously, it proposes a "local attention + graph convolution (GCN)" fusion mechanism based on the BARM module, achieving dynamic fusion through cross-feature enhancement, attention anchor construction, and global modeling of graph convolution, rather than static concatenation or weighting. These two technical approaches respectively solve the feature conflict problem of existing encoders and the modeling defects of the correlation of fusion modules, demonstrating originality in their technical approaches.

[0049] In summary, while existing technologies attempt to optimize boundary localization through edge supervision and multi-scale learning, none have achieved a systematic technological breakthrough at the three levels of labeling, architecture, and fusion. The technical means remain at the level of "supplementing edge information within the traditional framework." This invention, however, addresses the specific technical deficiencies of existing technologies by proposing a mathematical method for label decomposition, a symmetrical double-span structure for the encoder, and an attention-graph convolution combination mechanism for the fusion module. Each technical means addresses a specific technical pain point and forms a collaborative optimization technical system. Compared with existing technologies, it has fundamental advantages in boundary localization accuracy, feature extraction efficiency, and fusion rationality, while maintaining the model's lightweight nature (39.14M parameters), demonstrating greater technological innovation and practicality.

[0050] 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 label decoupling, characterized in that, Includes the following steps: The tampered image to be detected is acquired, and the original binary tampering mask corresponding to the tampered image is decomposed into a body image and a detail image through a label decoupling process; wherein, the body image is used to represent the main structure of the tampered region, and the detail image is used to represent the boundary details of the tampered region; The processed tampered image is input into a dual-span hourglass visual Transformer encoder for feature extraction, obtaining body feature streams and detail feature streams adapted to the body image and detail image, respectively. The body feature stream and detail feature stream are respectively input into two independent paths in the bidirectional decoupled pyramid decoder for parallel refinement and optimization, resulting in optimized body features and detail features. The optimized body features and detail features are input into the body-detail adaptive relationship module for adaptive fusion to generate a final prediction mask that combines structural integrity and boundary accuracy. A multi-task loss function is used to jointly supervise the body prediction corresponding to body features, the detail prediction corresponding to detail features, and the final prediction mask to ensure that the components of the model are optimized collaboratively.

2. The method according to claim 1, characterized in that, The tag decoupling process specifically includes: Decomposition is achieved by performing morphological operations on the original binary tampering mask, specifically by obtaining the body image through dilation and obtaining the detail image by performing a difference operation between the original binary tampering mask and the body image.

3. The method according to claim 1, characterized in that, The dual-span hourglass visual Transformer encoder adopts an hourglass-shaped structure, including a sequentially connected downsampling path and an upsampling path; The downsampling path is used to extract coarse-grained, high-semantic body feature streams; the upsampling path is used to extract fine-grained, high-resolution detail feature streams. The fusion and transfer of features in the downsampling path and the upsampling path are achieved through a double-span skip connection.

4. The method according to claim 1, characterized in that, The bidirectional decoupled pyramid decoder includes a body grouping fusion module and a detail grouping fusion module configured in parallel. The body grouping and fusion module is used to refine the body feature stream at multiple scales, capture the global structural information of the tampered region through multi-branch convolution operations, and output the optimized body features. The detail grouping and fusion module is used to refine the detail feature stream at multiple scales. It captures subtle traces of boundary tampering through multi-branch convolution operations and outputs optimized detail features.

5. The method according to claim 1, characterized in that, The adaptive fusion operation of the body-detail adaptive relationship module specifically includes the following: Local attention modeling is performed on the optimized body features and detail features of the input to highlight key regions related to tampering; The attention-modeled features are converted into query vectors, key vectors, and value vectors, and the association between the optimized body features and detail features is calculated through the attention mechanism. Graph convolutional networks are used to perform graph convolution operations on the associated features to model the global relationships between features. By fusing the features processed by graph convolution through multiplication and addition operations, a fused feature is obtained to generate the final tamper location prediction mask.

6. The method according to claim 1, characterized in that, The multi-task loss function includes a supervised loss for predicting the body branch, a supervised loss for predicting the detail branch, and a supervised loss for predicting the final prediction mask.

7. A tag-decoupled image tampering localization system for implementing the method of any one of claims 1-6, characterized in that, include: The image acquisition module is used to acquire the tampered image to be detected; The label decoupling module is used to decompose the original binary tampering mask corresponding to the tampered image into a body image and a detail image through the label decoupling process; The feature encoding module includes a dual-span hourglass visual Transformer encoder, which is used to extract features from tampered images and output body feature streams and detail feature streams. The feature decoding optimization module includes a bidirectional decoupled pyramid decoder, which is used to refine and optimize the body feature stream and detail feature stream in parallel through two independent paths to obtain optimized body features and detail features; The feature fusion module includes a body-detail adaptive relationship module, which is used to adaptively fuse optimized body features and detail features to generate the final tamper localization prediction mask; The training supervision module is used to jointly supervise the model based on the body prediction corresponding to the body features, the detail prediction corresponding to the detail features, and the final prediction mask, and then construct a multi-task loss function to jointly supervise the training of the model.

8. The system according to claim 1, characterized in that, The bidirectional decoupling pyramid decoder specifically includes: The body grouping and fusion submodule is used to refine the body feature flow in a multi-scale structure. The detail grouping and fusion submodule is used to refine the detail feature stream at multiple scales. The body grouping fusion submodule and the detail grouping fusion submodule are structurally independent and work in parallel.

9. 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 tag decoupling as described in any one of claims 1-6.

10. 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 tag decoupling as described in any one of claims 1-6.

Images