Image tampering positioning system and method based on compound eye bionics and uncertainty perception

Abstract

The invention discloses an image tampering positioning system and method based on compound eye bionics and uncertainty perception, and belongs to the technical field of image forensics and deep learning. The system provided by the invention comprises a multi-source parallel encoder which adopts images with three resolutions as input and retains tampering traces with different scales; the multi-resolution integration unit is used for dynamically fusing multi-resolution features through cross-scale attention; the chain type grouping interaction decoder is used for gradually eliminating redundant information and enhancing the discrimination capability of a tampered area; gaussian uncertainty loss is used for explicit penalty low-confidence prediction, and the positioning precision of the fuzzy boundary is improved. The method provided by the invention comprises the following steps: providing a bionic multi-resolution uncertainty sensing network, and realizing high-precision image tampering positioning by simulating a multi-aperture parallel sensing mechanism of an insect compound eye and combining multi-resolution input, chained grouping interactive decoding and Gaussian uncertainty loss.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image forensics and deep learning, and particularly relates to an image tampering positioning system and method based on compound eye bionics and uncertainty perception. BACKGROUND

[0002] Image Manipulation Localization (IML) is a key technology in digital image forensics, aiming to identify tampered regions in images through pixel-level positioning to ensure the authenticity of multimedia content. With the popularity of deepfake and image editing tools, IML has become increasingly important in news verification, judicial forensics, and social media content review.

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

[0004] Loss of multi-scale tampering traces: Tampering traces (such as inconsistent edges, abnormal noise patterns, etc.) exhibit different characteristics at different resolutions, but existing methods usually construct a feature pyramid from a single resolution image, resulting in the loss of high-resolution details during encoding, especially for subtle traces that are only significant at specific scales.

[0005] Prediction uncertainty in ambiguous regions: In tampering boundaries or texture complex regions, models are prone to produce low-confidence predictions due to information ambiguity, while existing loss functions (such as binary cross-entropy loss) treat all pixels equally, failing to optimize high-uncertainty regions.

[0006] Existing IML methods can be mainly divided into:

[0007] (1) Feature pyramid method based on single resolution input

[0008] Representative schemes: MVSS-Net (Chen et al., ICCV 2021), TruFor (Guillaro et al., CVPR 2023), etc. The technical details include:

[0009] Construct a multi-scale feature pyramid from a single resolution image (such as 512x512) through a convolutional neural network (CNN) or Transformer.

[0010] Generate feature maps of different scales through upsampling or downsampling, but high-resolution details are lost in multiple downsampling processes.

[0011] The limitations are:

[0012] The construction of feature pyramid relies on single input, leading to subtle tampering traces (such as local noise inconsistency) being diluted in low-resolution levels.

[0013] Insufficient positioning ability for ambiguous boundaries or complex texture regions, prediction results have uncertainty.

[0014] (2) Frequency domain enhancement method based on dual-stream architecture

[0015] Representative schemes: SPAN (Hu et al., ECCV 2020), EITLNet (Guo et al., ICASSP 2024). Technical details include:

[0016] Parallel processing of RGB spatial domain and frequency domain (such as DCT or high-pass filter features) to capture frequency domain anomalies introduced by tampering (such as JPEG compression traces).

[0017] Fusion of dual-stream features through attention mechanism.

[0018] Its limitations are:

[0019] Frequency domain features are not sensitive to certain tampering types (such as smooth area replication), and rely on manually designed filters.

[0020] No explicit modeling of prediction uncertainty, insufficient optimization for low confidence regions.

[0021] (3) Auxiliary loss method for uncertainty modeling

[0022] Representative schemes: MVSS-Net++ (introducing edge-aware loss), IML-ViT (Ma et al., AAAI 2024). Technical details include:

[0023] Increase edge supervision or contrast loss based on standard binary cross-entropy (BCE) loss to improve boundary positioning accuracy.

[0024] Its limitations are:

[0025] Dependence on explicit edge annotation, limited effect on tampering without clear boundaries (such as blur processing).

[0026] No modeling of pixel-level uncertainty from a probabilistic perspective, unable to dynamically adjust the weight of fuzzy regions.

[0027] The main shortcomings of existing technologies can be summarized as:

[0028] (1) Insufficient multi-scale information integration: single resolution input leads to loss of key details in feature pyramid.

[0029] (2) Redundancy information suppression missing: multi-source feature fusion does not effectively filter noise.

[0030] (3) Weakness in uncertainty processing: loss function does not explicitly optimize low confidence area.

[0031] The causes are as follows:

[0032] (1) The existing network generally adopts "binary cross entropy + Dice / IoU Loss" training, and the supervision signal is a hard 0 / 1 mask; the network is forced to give an absolute 0 or 1 decision for the edge, compression artifact or slightly post-processed area, and cannot output a "confidence map". Experiments show that slight JPEG re-compression can cause the predicted edge pixels to drift significantly, and further cause visual "jagged" or "halo" effects.

[0033] (2) Although the MVSS-Net inputs three resolutions of 256 / 512 / 768 in parallel, it only performs one-way fusion through the top-down FPN, and lacks cross-scale "error correction" or "consistency constraint". When the tampered area accounts for only 1% of the image, the low-resolution branch loses details; when the tampered area covers more than 50%, the high-resolution branch has insufficient receptive field, resulting in missed detection or over-detection.

[0034] (3) The existing method does not explicitly introduce "domain difference" prior in the training stage. The network tends to remember the JPEG quantization table, camera CFA mode or GAN generated texture specific to the training set; once a new tampering type is encountered during testing, the performance will decrease significantly.

[0035] In order to solve the above problems, the present application proposes an image tampering positioning system and method based on compound eye bionics and uncertainty perception. SUMMARY

[0036] The purpose of the present application is to propose an image tampering positioning system and method based on compound eye bionics and uncertainty perception to solve the problem that the prior art is difficult to guarantee positioning accuracy and model generalization ability at the same time when facing images with multiple resolutions and multiple tampering methods, and the uncertainty estimation of the tampered area is insufficient, resulting in blurred boundaries, false detection or missed detection in the detection result. The present application realizes accurate and robust positioning of the image tampered area by fusing different scale inputs, explicitly modeling uncertainty and optimizing deep supervision strategy.

[0037] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0038] The image tampering positioning system based on compound eye bionics and uncertainty perception comprises:

[0039] The multi-source parallel encoder adopts three resolutions of images as input, and retains different scale tampering traces;

[0040] Multi-resolution integration unit, dynamically fusing multi-resolution features through cross-scale attention;

[0041] Chain grouping interaction decoder, used to eliminate redundant information step by step and enhance the discrimination ability of tampered regions;

[0042] Gaussian uncertainty loss, used to explicitly punish low-confidence predictions and improve the positioning accuracy of fuzzy boundaries.

[0043] Preferably, the execution of the multi-source parallel encoder specifically includes the following contents:

[0044] Spatial context branch: input low, medium and high resolution images, generate hierarchical features {f0, f1, f2, f3} through PVTv2 encoder with shared weights;

[0045] High-frequency artifact branch: apply channel high-pass filter to medium resolution image, extract frequency domain features through lightweight MiT encoder

[0046] Preferably, the execution of the multi-resolution integration unit specifically includes the following contents:

[0047] Cross-scale feature enhancement: low→medium resolution fusion, used to enhance the local details of medium resolution features; medium→high resolution fusion, used to preserve global context information; the formula representation is as follows:

[0048] P1=u(f i low )

[0049]

[0050] In the formula, P1 represents up-sampling of low resolution features; P2 represents adding up-sampled P1 and medium resolution features and multiplying by 1 / 2; P3 represents adding P2 and high resolution features and then multiplying by 1 / 2;

[0051] Multi-head spatial attention fusion: concatenate P1, P2, P3 along the channel dimension, dynamically weighted through grouping attention mechanism; each group independently generates attention weights, highlighting key tampering traces; the formula representation is as follows:

[0052] F i =Cat[P1,P2,P3]

[0053]

[0054] In the formula, F i represents the fused feature descriptor; Cat represents concatenation along the channel; f iMRI representing spatial enhanced features; representing element-wise multiplication;

[0055] F i processed by M parallel attention groups; in each group m, two independent linear transformations τ i are projected into attention weights and feature representations respectively, the attention normalized by softmax is modulated to the corresponding features through element-wise multiplication , the weighted outputs of all groups are aggregated to generate the final spatial enhanced features f i MRI This multi-head design realizes comprehensive spatial interaction while maintaining computational efficiency.

[0056] Preferably, the execution of the chain group interaction decoder specifically includes the following contents:

[0057] Based on the chain group interaction decoder, multi-source feature aggregation is realized:

[0058] Input multi-resolution integrated unit features f MRI , frequency domain features f freq and upper layer decoder features

[0059] Chain group interaction is performed, and the mixed features are divided into n groups by channel, and each group is further divided into modulation set interaction set content set , wherein the interaction set is transferred from group k to group k+1 to realize cross-group information iterative optimization; the content set is concatenated and a gating vector is generated through a lightweight network to highlight the tampering traces;

[0060] Residual learning and output, the gating features are connected with the initial feature residual, and the output f i CGI .

[0061] Preferably, the execution of the Gaussian uncertainty loss specifically includes the following contents:

[0062] The prediction uncertainty is explicitly penalized by a Gaussian function, and the peak steepness of the penalty is controlled by a hyperparameter, and the specific formula is as follows:

[0063] L GUL = exp(-α(p i -0.5) 2 )

[0064] In the formula, L GULrepresents the Gaussian uncertainty loss, used to punish the uncertainty in the model prediction; alpha represents a hyperparameter, used to control the steepness of the punishment peak;

[0065] The Gaussian uncertainty loss is used to force the model to make a high confidence judgment on the ambiguous area, and the L bce are used together, and the hyperparameters are set according to experiments, and the specific formula is as follows:

[0066]

[0067] The image tampering positioning method based on compound eye bionics and uncertainty perception includes the following contents:

[0068] Based on the image tampering positioning system, a bionic multi-resolution uncertainty perception network is proposed, which simulates the multi-aperture parallel perception mechanism of insect compound eyes, combines multi-resolution input, chain grouping interactive decoding and Gaussian uncertainty loss, and realizes high-precision image tampering positioning.

[0069] The application further protects a computer device, which comprises a processor and a memory, and the memory stores at least one instruction, at least one program, a code set or an instruction set, which is loaded and executed by the processor to realize the above-mentioned image tampering positioning system based on compound eye bionics and uncertainty perception.

[0070] The application further protects a computer readable storage medium, which stores at least one instruction, at least one program, a code set or an instruction set, which is loaded and executed by the processor to realize the above-mentioned image tampering positioning system based on compound eye bionics and uncertainty perception.

[0071] Compared with the prior art, the application provides an image tampering positioning system and method based on compound eye bionics and uncertainty perception, which has the following beneficial effects:

[0072] (1) The application can retain multi-scale tampering details and reduce key information loss:

[0073] Compared with the prior art, the existing IML method usually constructs a feature pyramid from a single resolution image to obtain multi-scale information, which inevitably loses high-resolution tampering traces that are only obvious at a certain scale during the encoding process.

[0074] The application learns from the multi-aperture parallel imaging principle of insect compound eyes, introduces multi-resolution image streams in parallel at the input end, and dynamically cross-scale interacts and grouping attention fusion through a multi-resolution fusion unit (MRI), directly retains key tampering clues at different scales at the original signal level, thereby significantly improving the detection capability of fine-grained tampering areas.

[0075] (2) The present application can inhibit the redundant information introduced by multi-resolution, and enhance the feature discrimination:

[0076] Compared with the prior art, the existing multi-resolution or multi-branch method directly splices features after parallel decoding in the fusion stage, lacks a fine redundant suppression mechanism across scales and groups, and is prone to noise accumulation.

[0077] Although multi-resolution input can enrich information, it may also introduce noise and redundancy. The present application designs a chain grouping interaction unit (CGI), which groups channels and adopts a chain information transmission and dynamic gating mechanism in the decoding stage, to transmit context information step by step, suppress irrelevant features, and strengthen significant tampering traces, thereby achieving higher confidence positioning results.

[0078] (3) The present application can improve the discrimination ability in fuzzy areas by explicitly modeling the uncertainty:

[0079] Compared with the prior art, the existing method uses BCE to supervise all pixels equally, or relies on explicit edge information, and is difficult to deal with complex texture or fuzzy boundary areas.

[0080] The present application proposes a Gaussian uncertainty loss (L GUL ), which sets a strong penalty peak near the predicted value 0.5, explicitly amplifies the loss of high-uncertainty pixels, guides the model to make more decisive predictions in fuzzy areas, and can effectively reduce false positives and false negatives.

[0081] (4) The present application has strong cross-distribution generalization ability and can adapt to various tampering scenarios:

[0082] Compared with the prior art, the existing method mostly relies on a single modality (such as the spatial domain) or single-resolution features, resulting in a significant performance decline in unseen tampering types or distributions.

[0083] The present application simultaneously processes multi-resolution images and frequency domain features at the input end, adopts a multi-source parallel coding structure, and complements and fuses spatial domain multi-scale features and frequency domain high-frequency artifact information, so that the model learns the intrinsic tampering features independent of specific data distribution. Combined with the redundancy suppression mechanism of MRI and CGI and the uncertainty constraint, the present application can maintain stable high performance on different types of tampering and different distribution datasets, and significantly outperforms the prior art in the cross-dataset (OOD) scenario. BRIEF DESCRIPTION OF DRAWINGS

[0084] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and other accompanying drawings can be obtained by those skilled in the art without creative labor on the basis of these accompanying drawings.

[0085] Figure 1 The overall architecture diagram of the image tampering positioning system based on compound eye bionics and uncertainty perception proposed by the present application is shown in the figure.

[0086] Figure 2 The variation trend diagram of the Gaussian uncertainty loss with the model uncertainty proposed by the present application is shown in the figure. DETAILED DESCRIPTION

[0087] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all the embodiments.

[0088] The present application proposes an image tampering positioning system and method based on compound eye bionics and uncertainty perception. By introducing quantifiable uncertainty estimation, the problem of "edge blur and lack of confidence" is solved. By explicitly modeling pixel-level Gaussian uncertainty at the output end, the network gives an "uncertainty" prompt in the boundary or compressed, post-processing interference area, so that the user or downstream system can make a secondary decision based on the confidence map, reducing false positives and false negatives. In addition, the present application establishes cross-scale cooperation and consistency constraints to improve the adaptability to different sizes of tampering. On the basis of multi-resolution input, a cross-scale error correction module (MRI) is added to enable the 256 / 512 / 768 three branches to interact bidirectionally at the feature level and the output level, retaining high-resolution details and utilizing low-resolution global context to alleviate the contradiction between "difficult to detect small tampering and easy to detect large tampering". Further, the present application adopts dynamic depth supervision weight based on uncertainty, reduces the manual optimization of hyperparameters, introduces Gaussian uncertainty loss (L GUL ) into each intermediate supervision layer, and adjusts the loss weight in real time according to the prediction uncertainty of each layer and each sample, realizes adaptive training of "small weight for large noise and large weight for small noise", reduces the overfitting risk caused by fixed λ, and improves the cross-domain migration ability.

[0089] In summary, this invention aims to systematically improve existing methods in terms of edge accuracy, scale adaptability, hyperparameter robustness, and cross-domain generalization by focusing on "uncertainty perception" and "multi-resolution consistency," rather than negating their existing contributions. The image tampering localization system and method based on compound eye bionics and uncertainty perception proposed in this invention will be further described below with reference to relevant figures and specific examples.

[0090] Example 1:

[0091] This invention proposes an image tampering localization system and method based on compound eye bionics and uncertainty perception, specifically including the following:

[0092] A biomimetic multi-resolution uncertainty perception network (Omma-Net) is proposed. By simulating the multi-aperture parallel perception mechanism of insect compound eyes, and combining multi-resolution input, chain grouping interaction (CGI), and Gaussian uncertainty loss (GUL), high-precision image tampering localization is achieved. Specifically, it includes:

[0093] Multi-resolution parallel input: Images with three resolutions (such as 224×224, 512×512, and 768×768) are used as input to preserve tampering traces at different scales.

[0094] Multi-Resolution Integration (MRI): Dynamically fuses multi-resolution features through cross-scale attention.

[0095] Chained Grouped Interaction Unit (CGI): Gradually eliminates redundant information and enhances the ability to detect tampered areas. Gaussian Uncertainty Loss (GUL): Explicitly penalizes low-confidence predictions, improving the positioning accuracy of ambiguous boundaries.

[0096] More specifically, it includes the following:

[0097] 1. Multi-source parallel encoder:

[0098] Spatial context branch: Input three-resolution images (low, medium, high), and generate hierarchical features {f0, f1, f2, f3} through a PVTv2 encoder with shared weights;

[0099] High-frequency artifact branch: Apply a channel high-pass filter (CW-HPF) to the medium-resolution image and extract frequency domain features using a lightweight MiT encoder.

[0100] 2. Multi-resolution integrated unit (MRI), combined with Figure 1 (b) is mainly divided into two stages:

[0101] Cross-scale feature enhancement: Low-to-medium resolution fusion enhances local details of medium-resolution features. Medium-to-high resolution fusion preserves global contextual information.

[0102] P1=u(f i low )

[0103]

[0104] In the formula, P1 represents upsampling of low-resolution features; P2 represents adding the upsampled P1 and the medium-resolution features and multiplying by 1 / 2; P3 represents adding P2 and the high-resolution features and then multiplying by 1 / 2.

[0105] Multi-head spatial attention fusion: P1, P2, and P3 are concatenated along the channel dimension and dynamically weighted using a grouped attention mechanism. Each group generates its own attention weight, highlighting key tampering traces.

[0106] F i =Cat[P1,P2,P3]

[0107]

[0108] In the formula, F i This represents the fused feature descriptor; Cat indicates splicing along the channel; f i MRI Represents spatially enhanced features; This represents element-wise multiplication;

[0109] The fused feature descriptor F i Processing is performed through M parallel attention groups; in each group m, two independent linear transformations τ will F i The projections are respectively the attention weights and feature representations. The attention, after softmax normalization, is then processed through element-wise multiplication. The corresponding features are modulated, and the weighted outputs of all groups are aggregated to generate the final spatial augmentation feature f. i MRI .

[0110] 3. Chained Group Interactive Decoder (CGI) (combined with...) Figure 1 (c)):

[0111] Multi-source feature aggregation: Input multi-resolution integrated unit (MRI) features f MRI Frequency domain characteristics f freqand features of the upper-layer decoder Then, through chain-like grouping and interaction, the mixed features are divided into n groups according to channels, and each group is further divided into modulation sets. Interaction set Content Collection Interaction set Information is passed from group k to group k+1 for iterative optimization across groups. After concatenation of content sets, a lightweight network generates gated vectors, which are weighted to highlight tampering traces. Finally, residual learning and output are performed: the gated features are concatenated with the residuals of the initial features to output f. i CGI .

[0112] 4. Gaussian Uncertainty Loss (L) GUL The trend of change with the predicted value is as follows: Figure 2 As shown. GUL explicitly penalizes prediction uncertainty using a Gaussian function, with hyperparameter α = 8, controlling the steepness of the penalty peak ( Figure 2 The maximum loss is shown when the predicted value is 0.5.

[0113] L GUL =exp(-α(p) i -0.5) 2 )

[0114] In the formula, L GUL α represents the Gaussian uncertainty loss, used to penalize the uncertainty in the model's predictions; α represents the hyperparameter, used to control the steepness of the penalty peak; through extensive experiments, we determined that when α = 8, the model can achieve the best balance between sharpening the segmentation boundaries and maintaining training stability, thus achieving the best overall performance.

[0115] Function: To force the model to make high-confidence judgments on ambiguous regions (such as boundaries).

[0116] The Gaussian uncertainty loss in this application requires the cooperation of L. bce Use them together, and set the hyperparameters according to the experiment as shown in the following formula:

[0117]

[0118] This formulaic expression employs a deep supervision strategy, by applying L to four different decoder levels. bce The loss is calculated using experimentally determined weights. These weights are set to progressively increase the emphasis on shallower layers, which generate more refined feature maps. This design balances the need for pixel-level accuracy across all scales while also providing a strong impetus for generating a well-defined, unambiguous final localization mask.

[0119] Example 2:

[0120] Based on Example 1, but with a difference, the proposed image tampering localization system based on compound eye bionics and uncertainty perception was tested using multiple benchmark datasets. The specific test results are as follows:

[0121] Table 1

[0122]

[0123] Note: In the evaluation results of all the above comparison methods on the internal distribution (ID) dataset, the best performance is first indicated by bold, and the second best performance is indicated by underline.

[0124] Table 2

[0125]

[0126] Note: In the evaluation results of all the above comparison methods on the Outlier Distribution (OOD) dataset, the best performance is indicated by bold first, and the second best performance is indicated by underline.

[0127] Table 3

[0128]

[0129] As can be clearly seen from the above tables, the present invention achieves superior positioning accuracy, generalization and robustness compared to existing technologies on nine benchmark datasets.

[0130] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An image tampering localization system based on compound eye bionics and uncertainty perception, characterized in that, The system includes: The multi-source parallel encoder uses images of three resolutions as input to preserve tampering traces at different scales; Multi-resolution fusion unit dynamically fuses multi-resolution features through cross-scale attention; A chained group interactive decoder is used to gradually eliminate redundant information and enhance the ability to detect tampered areas; Gaussian uncertainty loss is used to explicitly penalize low-confidence predictions and improve the positioning accuracy of fuzzy boundaries.

2. The image tampering localization system based on compound eye bionics and uncertainty perception according to claim 1, characterized in that, The execution of the multi-source parallel encoder specifically includes the following: Spatial context branch: Input images of low, medium and high resolutions, and generate hierarchical features {f0, f1, f2, f3} through a PVTv2 encoder with shared weights; High-frequency artifact branch: Apply a channel high-pass filter to the medium-resolution image and extract frequency domain features using a lightweight MiT encoder.

3. The image tampering localization system based on compound eye bionics and uncertainty perception according to claim 2, characterized in that, The execution of the multi-resolution integration unit specifically includes the following: Cross-scale feature enhancement: Low-to-medium resolution fusion is used to enhance local details of medium-resolution features; medium-to-high resolution fusion is used to preserve global contextual information; the formulas for the above are expressed as follows: P1=u(f i low ) In the formula, P1 represents upsampling of low-resolution features; P2 means adding the upsampled P1 and the medium-resolution feature and multiplying by 1 / 2; P3 means adding P2 and the high-resolution feature and then multiplying by 1 / 2. Multi-head spatial attention fusion: P1, P2, and P3 are concatenated along the channel dimension and dynamically weighted through a grouped attention mechanism; each group generates its own attention weight to highlight key tampering traces; the formula for the above is as follows: F i =Cat[P1,P2,P3] In the formula, F i This represents the fused feature descriptor; Cat indicates splicing along the channel; f i MRI Represents spatially enhanced features; This represents element-wise multiplication; The fused feature descriptor F i Processing is performed through M parallel attention groups; in each group m, two independent linear transformations τ will F i The attention weights and feature representations are projected separately, and the attention weights, after being normalized by softmax, are multiplied element-wise. The corresponding features are modulated, and the weighted outputs of all groups are aggregated to generate the final spatial augmentation feature f. i MRI .

4. The image tampering localization system based on compound eye bionics and uncertainty perception according to claim 3, characterized in that, The execution of the chained group interactive decoder specifically includes the following: Multi-source feature aggregation based on chained group interactive decoder: Input multi-resolution ensemble unit features f MRI Frequency domain characteristics f freq and features of the upper-layer decoder A chain-like grouping interaction is performed, dividing the mixed features into n groups according to channels, and each group is further divided into modulation sets. Interaction set Content Collection Among them, the interaction set Information is passed from group k to group k+1 to achieve cross-group information iterative optimization; content set After concatenation, a lightweight network is used to generate gated vectors, which are then weighted to highlight tampering traces. Residual learning and output, gated features and initial feature residuals are concatenated, and the output f is obtained. i CGI .

5. The image tampering localization system based on compound eye bionics and uncertainty perception according to claim 4, characterized in that, The execution of the Gaussian uncertainty loss specifically includes the following: Uncertainty in prediction is explicitly penalized using a Gaussian function, and the steepness of the penalty peak is controlled by hyperparameters. The specific formula is as follows: 50 GUL =exp(-α(p i -0.5) 2 ) In the formula, L GUL represents the Gaussian uncertainty loss, used to penalize the uncertainty in the model's predictions; α represents the hyperparameter, used to control the steepness of the penalty peak. Using a Gaussian uncertainty loss forced model to make high-confidence judgments on fuzzy regions, further combined with L... bce Use them together and set the hyperparameters according to the experiment. The specific formula is as follows:

6. The image tampering localization method based on compound eye bionics and uncertainty perception applied in the system as described in any one of claims 1-5, characterized in that, Includes the following: Based on the image tampering localization system, a biomimetic multi-resolution uncertainty perception network is proposed. By simulating the multi-aperture parallel perception mechanism of insect compound eyes, and combining multi-resolution input, chain-based group interactive decoding, and Gaussian uncertainty loss, high-precision image tampering localization is achieved.

7. 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 localization system based on compound eye bionics and uncertainty perception as described in any one of claims 1-5.

8. 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 localization system based on compound eye bionics and uncertainty perception as described in any one of claims 1-5.

Images