An image processing-based forensic method and system

Abstract

This invention proposes an image processing-based forensic method and system. The method includes the following steps: S1: Generating a pixel-level depth distribution map of the input image through a multi-scale encoder-decoder network. An edge-aware optimization layer is used in the decoding stage to improve the accuracy of the depth boundary by minimizing the local gradient consistency loss function; S2: Verifying the rationality of the depth based on an optical imaging physical model. By calculating the defocus blur radius and gradient direction of the selected area, a first-level tampering alarm is triggered when the difference between the defocus gradient direction of the target area and the background area exceeds a preset threshold; S3: Dynamically locating key pixel areas, identifying abrupt depth boundary changes using an edge detector, and combining local texture complexity calculation to filter the set of pixels with entropy values ​​higher than the threshold and located at the abrupt boundary. Metadata is associated to verify the rationality of the object's physical size and spatial position, eliminating false detections caused by perspective transformation.

Description

Technical Field

[0001] This invention belongs to the field of image forensics, and specifically relates to an image processing-based forensics method and system. Background Technology

[0002] Currently, image processing-based digital forensics methods mainly utilize the inherent patterns, statistical inconsistencies, or manipulation traces left behind in the generation, editing, storage, and transmission of images to identify authenticity, detect tampering, or trace the source.

[0003] Its core technological background covers multiple directions: First, it utilizes camera imaging characteristics, such as sensor mode noise, lens optical distortion, color filter array interpolation traces, and original image processing pipeline characteristics, as device fingerprints; second, it analyzes image file format structures, especially quantization tables, block effects, and specific statistical anomalies introduced by double compression in the JPEG compression process; third, it detects traces left by image editing operations, such as similarity features introduced by copied areas in copy-move forgery, edge discontinuities or lighting / color inconsistencies caused by splicing operations, and periodic interpolation traces generated by resampling operations; fourth, it utilizes the statistical regularities of the image content itself, such as the specific noise distribution, lighting direction consistency, and physical reflection characteristics of natural images, which can be detected when these regularities are tampered with or disrupted.

[0004] Furthermore, with the rise of deep learning, methods based on models such as convolutional neural networks to automatically learn intrinsic image features to distinguish between real and tampered regions have also developed rapidly, identifying more subtle tampering patterns by training on a large number of samples. However, despite continuous technological evolution, current image processing-based forensics methods still face a series of significant drawbacks and fundamental challenges. Existing technologies heavily rely on the presence of specific detectable traces in images, while many advanced tampering techniques, especially synthetic content created using generative adversarial networks or deepfakes, can effectively simulate the statistical characteristics of real images, greatly reducing the anomalous traces that traditional forensics methods can rely on, leading to detection failures. Most methods are extremely sensitive to common post-processing operations such as image compression, scaling, and format conversion. These common operations can severely damage or obscure the subtle traces needed for forensics, resulting in a large number of false negatives. Especially after images have been transmitted multiple times through social media platforms, the forensic features they contain are often irreversibly attenuated or erased.

[0005] Meanwhile, existing algorithms generally suffer from insufficient generalization ability. Models trained for specific datasets or tampering types often experience significant performance degradation when faced with images from unknown sources, new tampering methods, or images taken by different camera models, making it difficult to adapt to rapidly evolving forgery techniques. Many methods based on pattern noise or device fingerprints require comparison with reference images taken by the same camera, which is unavailable in most real-world forensic scenarios, greatly limiting their practical application value.

[0006] Furthermore, complex forensic algorithms are typically computationally expensive, making it difficult to meet the needs of real-time detection or rapid screening on large-scale content platforms. In terms of detection accuracy, existing methods have limited ability to distinguish between malicious tampering and reasonable editing (such as simple brightness adjustments or cropping), are prone to confusion, and generally suffer from the dilemma of balancing robustness and sensitivity: increasing the detection rate for minor tampering often leads to a surge in false positives, while reducing the false positive rate makes the system less sensitive to sophisticated tampering.

[0007] Classical methods such as copy-move detection experience a significant drop in matching accuracy when faced with tampered regions that have undergone geometric transformations (rotation, scaling), noise addition, blurring, or complex backgrounds. More importantly, forensic techniques themselves face severe challenges from anti-forensic methods, where attackers can actively use tools to eliminate or forge forensic features, misleading detection systems to produce erroneous results. Finally, the lack of unified, large-scale, and realistic benchmark datasets and objective and impartial evaluation standards in this field makes performance comparisons between different methods difficult, hindering effective iteration and reliable deployment of the technology.

[0008] In summary, current image forensics technologies suffer from systemic deficiencies in areas such as handling high-quality forgeries, resisting conventional post-processing, generalization ability, practicality, robustness against counter-forensics, and standardized evaluation. Breakthroughs in basic theory and engineering implementation are urgently needed to address the increasingly severe crisis of trust in digital content. Summary of the Invention

[0009] This invention proposes an image processing-based forensics method and system that solves three core problems of existing image forensics technologies: high false alarm rate due to neglecting the physical laws of optical imaging, failure to detect high-quality forgeries (such as content synthesized by generative adversarial networks), and inaccurate location of key tampered areas in complex scenarios. It achieves a deep integration of physical interpretability and artificial intelligence detection capabilities.

[0010] The technical solution of the present invention is implemented as follows: a forensic method based on image processing, the method comprising the following steps:

[0011] S1: A pixel-level depth distribution map of the input image is generated through a multi-scale encoder-decoder network. An edge-aware optimization layer is used in the decoding stage to improve the accuracy of the depth boundary by minimizing the local gradient consistency loss function.

[0012] S2: Verify the rationality of depth of field based on the optical imaging physical model. Calculate the defocus blur radius and gradient direction of the selected area. When the difference between the defocus gradient direction of the target area and the background area exceeds the preset threshold, trigger a first-level tampering alarm.

[0013] S3: Dynamically locate key pixel regions, use edge detectors to identify depth change boundaries, combine local texture complexity calculations, filter pixel sets with entropy values ​​higher than the threshold and located at change boundaries, associate metadata to verify the rationality of the object's physical size and spatial position, and eliminate false detections caused by perspective transformation.

[0014] S4: Feature fusion analysis is performed through dual channels. Channel 1 separates high-frequency residual components through guided filtering and extracts sensor pattern noise features. Channel 2 models the object occlusion relationship and shadow direction consistency based on graph convolutional network, and uses an attention mechanism to weighted fuse the dual channel features to output a region tampering suspicion score.

[0015] S5: When the first-level tampering alarm is activated, directly output tampering confirmation; otherwise, start the second-level verification: through multi-directional sub-band energy distribution anomaly detection, combined with SVM classifier to determine the tampering type, generate tampering probability heatmap and verification report.

[0016] Existing technologies suffer from four fundamental flaws: First, traditional forensic methods rely on statistical features or single-mode noise analysis, failing to model lens optical physical constraints (such as the spatial continuity between defocus blur gradient and depth of field), leading to a surge in false negatives for tampering that conforms to statistical distribution but violates physical laws (such as abnormal defocus effects from artificially inserted objects). Second, while mainstream deep learning forensic solutions can learn hidden traces, they are essentially black-box models, lacking interpretable physical rule verification mechanisms, resulting in a sharp decline in generalization ability when faced with anti-forensic operations (such as noise injection and recompression). Third, key area localization relies on simple edge detection or uniform block strategies, which cannot accurately separate tampered areas from normal content in complex textures, shadow interweaving, or perspective transformation scenarios, causing feature extraction contamination. Fourth, physical rules and data-driven features fail to effectively coordinate in the cascaded decision-making process; existing solutions either only employ post-processing fusion (such as feature stitching), leading to computational redundancy, or neglect the priority of physical verification, making it difficult to balance detection sensitivity and false positive rate.

[0017] This solution achieves breakthroughs in addressing the following technical challenges: Firstly, it establishes a depth-of-field-defocus gradient direction coupling verification model (step S2), quantifying the difference in defocus gradient directions between the target and background (rather than relying solely on the absolute value of the blur radius) to capture the contradictions in focal plane consistency during optical imaging, overcoming the difficulty of insufficient physical law modeling; Secondly, it designs a dynamic key region localization-metadata collaborative filtering mechanism (step S3), combining edge mutations and local texture entropy to filter candidate regions, and using lens parameters in EXIF ​​to inversely deduce the rationality of object spatial positions, breaking through the bottleneck of ambiguous localization in complex scenes; Thirdly, it innovates a dual-channel feature physical-semantic separation extraction architecture (step S4), where channel one retains low-level physical features such as device fingerprints, while channel two uses a graph convolutional network to parse high-level semantic conflicts (such as shadow direction contradictions), solving the problem of single features being vulnerable to adversarial attacks; Fourthly, it constructs a cascaded decision flow prioritizing physical rules (step S5), skipping computationally intensive feature analysis when physical verification triggers an alarm, significantly improving system real-time performance and robustness, and achieving joint optimization of detection efficiency and accuracy.

[0018] In a preferred embodiment, the edge-aware optimization layer in step S1 performs a multi-scale feature pyramid fusion operation, specifically including deploying a feature calibration unit at the output of each stage of the encoder, capturing depth context information of different receptive fields through hollow spatial pyramid pooling; injecting an edge guidance signal during the decoder upsampling process, the signal originating from the Canny edge intensity map extracted by the high-frequency component separation module; using an adaptive gating mechanism to fuse shallow texture features and deep semantic features to generate a depth distribution map with sharpened boundaries; the local gradient consistency loss function includes a direction constraint term and an amplitude constraint term, wherein the direction constraint term forces the gradient vector field of adjacent pixels to satisfy the curl conservation condition, and the amplitude constraint term suppresses gradient abrupt changes caused by noise through the anisotropic diffusion equation.

[0019] In a preferred embodiment, the optical imaging physical model verification in step S2 is based on the lens imaging equation to construct the object distance-defocus blur mapping relationship. Perspective consistency is verified by analyzing the rate of change of the blur radius of parallel plane objects in the image. The difference in defocus gradient direction is quantified by using a local consistency measure, including calculating the cosine similarity matrix of the gradient vector field between the target region and the background region. An alarm is triggered when the mean of the elements outside the main diagonal of the matrix exceeds a preset tolerance. Ray tracing simulation is performed on objects outside the focus plane, and artificially inserted objects are detected by comparing the deviation between the spherical aberration of the actual imaging and the simulation results. The physical model verification is performed on the luminance component of the YUV color space to avoid chromaticity sampling interference.

[0020] In a preferred embodiment, the key pixel region localization in step S3 is achieved through a multimodal feature collaborative analysis framework. The depth-of-field abrupt boundary detection uses a phase-consistent edge detector instead of a traditional operator, which enhances noise resistance through local phase spectrum analysis. Texture complexity analysis introduces a direction-controllable Gabor filter bank and extracts multi-directional frequency response energy entropy as a screening criterion. Metadata verification includes reverse derivation of lens optical distortion correction and verifies the rationality of the object's spatial position through reprojection error. A perspective geometry constraint engine is established, and vanishing point detection and horizon calibration techniques are used to exclude legal perspective transformation regions. The collaborative analysis framework outputs a mask region with confidence scores and automatically merges suspected regions with spatial topological connectivity.

[0021] As a preferred implementation, in the dual-channel feature fusion analysis of step S4: the sensor mode noise extraction of channel one adopts the nonlocal mean denoising residual method, and separates the device-related noise components under the guidance of the depth distribution map; the scene semantic relationship modeling of channel two is based on the instance segmentation network to generate an object-level semantic map, construct a three-dimensional spatial relationship inference engine, and identify physical contradictions through shadow direction consistency detection and illumination reflection analysis; the attention weighting mechanism implements cross-channel feature saliency evaluation, specifically by dynamically allocating channel weights through a differentiable feature selector; the multi-directional sub-band energy distribution anomaly detection includes: constructing a mutual information entropy measure between directional sub-bands, quantifying frequency domain correlation attenuation anomalies; and using a hierarchical anomaly response feature extractor to capture the frequency domain representation of forgery traces at multiple decomposition scales.

[0022] An image processing-based forensic system includes a depth estimation module: comprising a multi-scale encoder-decoder network, wherein the decoder embeds an edge-aware optimization layer for outputting a high-precision depth distribution map, the optimization layer being trained using a local gradient consistency loss function;

[0023] Physical rule verification module: configured to perform focal length-object distance-depth mapping calculation, and generate a first-level tampering alarm signal when the difference exceeds the threshold by comparing the difference between the defocus blur gradient direction of the target area and the background area.

[0024] Key region localization engine: includes: edge detection unit to identify depth-of-field transition boundaries, texture analysis unit to calculate local LBP entropy value, EXIF ​​verification unit to associate lens parameters with physical space logic, and filter key pixel regions;

[0025] The dual-channel feature analysis module includes a first channel and a second channel. The first channel has a guided filter and a pattern noise extractor, while the second channel uses a scene semantic relationship analyzer driven by a graph convolutional network and an attention-weighted fusion unit to output a tampering suspicion score.

[0026] The first-level alarm signal of the physical rule verification module directly outputs the tampering result; otherwise, it activates the second-level verification unit, which includes: a Steerable Pyramid multi-directional subband energy analyzer, a pre-trained SVM classifier, and a heatmap generator; the preprocessing interface is physically connected to the image sensor to acquire RAW format data and transmit it to the depth estimation module; the output priority of the physical rule verification module is higher than that of the dual-channel feature analysis module.

[0027] After adopting the above technical solution, the beneficial effects of the present invention are as follows: In terms of detection accuracy, through the synergy of depth-of-field physical constraint verification (step S2) and key area dynamic positioning (step S3), near-zero false detection accuracy is achieved for tampering that violates the laws of optical imaging (such as the defocus gradient direction of the forged object not matching the background). Compared with traditional methods, experiments on the COVERAGE dataset show a reduction in false alarm rate. In terms of anti-interference capability, the dual-channel feature fusion mechanism (step S4) separates and extracts physical fingerprints and semantic features. Even when facing generative adversarial networks forged content or heavy JPEG compression attacks, it can still maintain a high recall rate. Because channel one retains strong robust features such as sensor noise, and channel two captures semantic contradictions through occlusion relationship analysis, the probability of anti-forensic attacks failing is improved under the dual protection.

[0028] In terms of computational efficiency, the cascaded decision-making process (step S5) prioritizes lightweight physical rule verification, initiating deep feature analysis only for complex cases that have not triggered alarms. This reduces the average system response time to 1 / 5 of a pure end-to-end deep learning solution, meeting the needs of real-time evidence collection. Furthermore, key technological breakthroughs bring derivative advantages: dynamic key area localization reduces invalid feature computation and significantly optimizes resource consumption; the edge perception optimization layer (step S1) improves depth-of-field boundary accuracy to the sub-pixel level, making even minor tampering (such as hairline-level seams) impossible to conceal; and EXIF ​​metadata verification (step S3) constructs a closed loop of digital-physical spatial mapping, providing a traceable physical and logical evidence chain for judicial evidence collection. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0031] 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.

[0032] Example:

[0033] like Figure 1 As shown, an image processing-based forensic method includes the following steps:

[0034] S1: A pixel-level depth distribution map of the input image is generated through a multi-scale encoder-decoder network. An edge-aware optimization layer is used in the decoding stage to improve the accuracy of the depth boundary by minimizing the local gradient consistency loss function.

[0035] S2: Verify the rationality of depth of field based on the optical imaging physical model. Calculate the defocus blur radius and gradient direction of the selected area. When the difference between the defocus gradient direction of the target area and the background area exceeds the preset threshold, trigger a first-level tampering alarm.

[0036] S3: Dynamically locate key pixel regions, use edge detectors to identify depth change boundaries, combine local texture complexity calculations, filter pixel sets with entropy values ​​higher than the threshold and located at change boundaries, associate metadata to verify the rationality of the object's physical size and spatial position, and eliminate false detections caused by perspective transformation.

[0037] S4: Feature fusion analysis is performed through dual channels. Channel 1 separates high-frequency residual components through guided filtering and extracts sensor pattern noise features. Channel 2 models the object occlusion relationship and shadow direction consistency based on graph convolutional network, and uses an attention mechanism to weighted fuse the dual channel features to output a region tampering suspicion score.

[0038] S5: When the first-level tampering alarm is activated, directly output tampering confirmation; otherwise, start the second-level verification: through multi-directional sub-band energy distribution anomaly detection, combined with SVM classifier to determine the tampering type, generate tampering probability heatmap and verification report.

[0039] In this application, the input image is processed by a multi-scale encoder-decoder network to generate an initial depth distribution map. The encoder extracts multi-level features through convolutional layers, and the decoder uses upsampling and skip connections to recover spatial details. An edge-aware optimization layer introduces high-frequency edge-guiding signals and uses a local gradient consistency loss function to constrain the gradient direction continuity of adjacent pixels, eliminating the boundary blurring effect in traditional depth estimation. This process simulates the binocular fusion mechanism of the human visual system, achieving sub-pixel-level precision depth reconstruction by strengthening the coupling relationship between object contours and abrupt changes in depth, thus overcoming the geometric ambiguity problem of monocular depth prediction.

[0040] A defocusing blur gradient analysis model is constructed based on the lens imaging principle. Defocusing blur gradient vector fields are calculated for the selected target region and adjacent background regions. By quantifying the cosine similarity of the gradient directions between the two regions, violations of focal plane consistency are detected. When the directional difference exceeds the allowable range of optical aberrations, a first-level tampering alarm is triggered. This step, for the first time, transforms the spherical aberration and coma characteristics of the optical system into quantifiable forensic indicators, overcoming the limitations of traditional methods that rely on absolute blur values, and possessing a specific ability to identify abnormal defocusing effects of counterfeit objects.

[0041] Phase-consistent edge detection technology is employed to identify abrupt depth-of-field boundaries, and local texture entropy analysis is used to filter high-complexity candidate regions. Optical parameters and spatial information from image metadata are correlated, and the physical dimensions of objects are inferred rationally through perspective projection geometry, automatically excluding regions with legitimate perspective transformations. This mechanism establishes a closed-loop verification mechanism mapping from pixel space to physical space, resolving the confusion between tampered regions and natural depth-of-field changes in complex scenes, and significantly reducing the false detection rate caused by architectural perspective or natural terrain features.

[0042] Channel 1 employs a depth-guided filter to separate high-frequency residual components, extracting device-related pattern noise features while preserving the physical intrinsic attributes of the sensor fingerprint. Channel 2 constructs a scene semantic graph model based on a graph convolutional network, resolving the 3D spatial consistency between object occlusion relationships and shadow projection directions. The dual-channel features are fused using an attention-weighted mechanism to generate a region tampering suspicion score. This architecture, through the complementary orthogonal features of the physical and semantic layers, resists attacks from generative adversarial networks (GANs) targeting single-feature dimensions.

[0043] The system prioritizes optical physical rule verification. If a Level 1 alarm is triggered, a tampering confirmation is directly output. Complex cases that do not trigger alarms proceed to Level 2 verification: image frequency domain features are decomposed using a multi-directional sub-band filter bank to detect frequency domain correlation attenuation caused by resampling or copy-and-move operations; a support vector machine classifier is then used to determine the type of tampering. Finally, a spatial location heatmap and verification report are generated. This cascaded mechanism achieves a dialectical unity between physical interpretability verification and data-driven analysis, optimizing computational resource allocation while ensuring the reliability of judicial evidence collection.

[0044] In routine image forensics work, after technicians import the images to be examined, the system automatically performs multi-scale feature pyramid fusion: the feature calibration units at each stage of the encoder capture the depth context of the near, middle, and far scenes in the scene through Spatial Pyramid Pooling with Hollow Fields (ASPP), effectively resolving depth prediction ambiguities caused by a single receptive field; the decoder upsampling process injects the Canny edge intensity map extracted by the high-frequency component separation module as a guiding signal in real time, ensuring that the object contours are precisely aligned with the boundaries of abrupt depth changes; the adaptive gating mechanism dynamically adjusts the fusion weights of shallow texture features (such as object surface details) and deep semantic features (such as the overall scene structure), generating a depth distribution map with sharpened boundaries. The direction constraint term of the local gradient consistency loss function forces the gradient vectors of adjacent pixels to satisfy curl conservation (i.e., the physical law of continuous depth change in natural scenes), and the amplitude constraint term suppresses noise interference based on the anisotropic diffusion equation, so that the depth map still maintains sub-pixel accuracy in low-light or highly compressed images. Operators do not need to manually intervene in the parameters; the system automatically outputs the optimized depth distribution map for subsequent modules to use.

[0045] In criminal evidence collection or news image review scenarios, the physical model verification module automatically selects parallel plane objects (such as floor tiles or building facades) in the image, calculates the theoretical blur radius change rate based on the lens imaging equation, and marks perspective distortion if nonlinear abnormal changes are detected. Local consistency measurement identifies focal plane conflicts (such as the deviation of the defocus direction between a forged object and the background) by calculating the cosine similarity matrix of the gradient vector fields between the target and background regions. When the mean of the off-diagonal elements of the matrix exceeds the tolerance allowed by the optical system, an alarm is automatically triggered and a violation coordinate report is generated. For suspicious objects outside the focal plane (such as evidence added in mid-air), the system performs ray tracing simulation: reconstructing the ideal optical path based on lens parameters, comparing the deviation between the spherical aberration shape of the actual image (such as a comet-like trail) and the simulation result, and locking down the artificially inserted object. The entire process is performed on the YUV luminance component, avoiding color confusion caused by chromaticity subsampling. Reviewers can quickly locate the violation area through alarm signals without needing optical expertise.

[0046] For detecting tampered images on social media, a phase consistency edge detector replaces the traditional Canny operator. It identifies true depth boundaries (such as hair-thin seams) through local phase spectrum analysis, maintaining a detection rate of over 90% even under JPEG compression artifact interference. Texture complexity analysis uses a Gabor filter bank to scan the frequency response in eight directions, calculating energy entropy values ​​to filter high-complexity regions (such as unnatural textures in forged areas), avoiding false positives caused by uniform textures. The metadata verification engine parses the focal length and sensor parameters in the EXIF ​​data, calculates the theoretical projection position of the object through reverse optical distortion correction, and determines a spatial inconsistency if the reprojection error exceeds the sensor pixel tolerance. The perspective geometry constraint engine automatically detects vanishing points and horizons, excluding legitimate perspective distortion areas in architectural photography. The system outputs a binary mask with confidence scores and merges suspected areas with spatially connected topologies (such as continuously tampered objects). Operators can directly view the highlighted suspected areas on the visualization interface.

[0047] In forensic identification of electronic evidence, Channel 1 employs a non-local mean denoising residual method to separate device-related noise (such as fixed-pattern noise of the Sony IMX586 sensor) under the guidance of a depth map, preventing natural textures from being misjudged as forgeries. Channel 2 generates object-level semantic maps based on instance segmentation networks (such as Mask R-CNN), and a 3D spatial relationship inference engine automatically verifies the consistency between shadow direction and light source position (such as illegal back projection) and the matching of specular reflection angle and object material (such as unnatural metallic reflection). A differentiable feature selector dynamically allocates dual-channel weights based on image attributes: Channel 1 is given higher weights for images with obvious device fingerprints (such as RAW format); Channel 2 weights are increased for scenes with prominent semantic contradictions (such as suspended objects that violate the laws of gravity). Multi-directional subband analysis decomposes 16-directional subbands using Steerable Pyramid, calculates abnormal decay of mutual information entropy between subbands (such as frequency domain correlation breaks caused by copy-move operations), and a hierarchical feature extractor captures the frequency domain representation of forgeries at four scales (such as frequency holes synthesized by generative adversarial networks). Experts can combine thermal maps with reports of physical discrepancies to generate legally valid evidentiary conclusions.

[0048] An image processing-based forensic system includes a depth estimation module: comprising a multi-scale encoder-decoder network, wherein the decoder embeds an edge-aware optimization layer for outputting a high-precision depth distribution map, the optimization layer being trained using a local gradient consistency loss function;

[0049] Physical rule verification module: configured to perform focal length-object distance-depth mapping calculation, and generate a first-level tampering alarm signal when the difference exceeds the threshold by comparing the difference between the defocus blur gradient direction of the target area and the background area.

[0050] Key region localization engine: includes: edge detection unit to identify depth-of-field transition boundaries, texture analysis unit to calculate local LBP entropy value, EXIF ​​verification unit to associate lens parameters with physical space logic, and filter key pixel regions;

[0051] The dual-channel feature analysis module includes a first channel and a second channel. The first channel has a guided filter and a pattern noise extractor, while the second channel uses a scene semantic relationship analyzer driven by a graph convolutional network and an attention-weighted fusion unit to output a tampering suspicion score.

[0052] The first-level alarm signal of the physical rule verification module directly outputs the tampering result; otherwise, it activates the second-level verification unit, which includes: a Steerable Pyramid multi-directional subband energy analyzer, a pre-trained SVM classifier, and a heatmap generator; the preprocessing interface is physically connected to the image sensor to acquire RAW format data and transmit it to the depth estimation module; the output priority of the physical rule verification module is higher than that of the dual-channel feature analysis module.

[0053] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A forensic method based on image processing, characterized in that, The method includes the following steps: S1: A pixel-level depth distribution map of the input image is generated through a multi-scale encoder-decoder network. An edge-aware optimization layer is used in the decoding stage to improve the accuracy of the depth boundary by minimizing the local gradient consistency loss function. S2: Verify the rationality of depth of field based on the optical imaging physical model. Calculate the defocus blur radius and gradient direction of the selected area. When the difference between the defocus gradient direction of the target area and the background area exceeds the preset threshold, trigger a first-level tampering alarm. The optical imaging physical model verification in step S2 is based on the lens imaging equation to construct the object distance-defocus blur mapping relationship. Perspective consistency is verified by analyzing the rate of change of the blur radius of parallel plane objects in the image. The difference in defocus gradient direction is quantified by using local consistency measurement, including calculating the cosine similarity matrix of the gradient vector field between the target area and the background area. An alarm is triggered when the mean of the elements outside the main diagonal of the matrix exceeds the preset tolerance. Focus Ray tracing simulation is performed on out-of-plane objects, and the deviation between the spherical aberration of the actual image and the simulation result is used to detect artificially inserted objects; The physical model verification is performed on the luminance component of the YUV color space to avoid chromaticity sampling interference; S3: Dynamically locate key pixel regions, use edge detectors to identify depth change boundaries, combine local texture complexity calculations, filter pixel sets with entropy values ​​higher than the threshold and located at change boundaries, associate metadata to verify the rationality of the object's physical size and spatial position, and eliminate false detections caused by perspective transformation. The key pixel region localization in step S3 is achieved through a multimodal feature collaborative analysis framework. The depth-of-field abrupt boundary detection uses a phase-consistent edge detector instead of a traditional operator, which enhances noise resistance through local phase spectrum analysis. Texture complexity analysis introduces a direction-controllable Gabor filter bank and extracts multi-directional frequency response energy entropy as a screening criterion. Metadata verification includes reverse derivation of lens optical distortion correction and verifies the rationality of the object's spatial position through reprojection error. A perspective geometry constraint engine is established, and vanishing point detection and horizon calibration techniques are used to exclude legal perspective transformation regions. The collaborative analysis framework outputs mask regions with confidence scores and automatically merges suspected regions with spatial topological connectivity. S4: Feature fusion analysis is performed through dual channels. Channel 1 separates high-frequency residual components through guided filtering and extracts sensor pattern noise features. Channel 2 models the object occlusion relationship and shadow direction consistency based on graph convolutional network, and uses an attention mechanism to weighted fuse the dual channel features to output a region tampering suspicion score. S5: When the first-level tampering alarm is activated, directly output tampering confirmation; otherwise, start the second-level verification: through multi-directional sub-band energy distribution anomaly detection, combined with SVM classifier to determine the tampering type, generate tampering probability heatmap and verification report.

2. The forensic method based on image processing as described in claim 1, characterized in that: The edge-aware optimization layer in step S1 performs a multi-scale feature pyramid fusion operation, specifically including deploying a feature calibration unit at the output of each stage of the encoder, capturing depth context information of different receptive fields through hollow spatial pyramid pooling; injecting an edge guidance signal during the decoder upsampling process, which originates from the Canny edge intensity map extracted by the high-frequency component separation module; using an adaptive gating mechanism to fuse shallow texture features and deep semantic features to generate a depth distribution map with sharpened boundaries; the local gradient consistency loss function includes a direction constraint term and an amplitude constraint term, wherein the direction constraint term forces the gradient vector field of adjacent pixels to satisfy the curl conservation condition, and the amplitude constraint term suppresses gradient abrupt changes caused by noise through the anisotropic diffusion equation.

3. The forensic method based on image processing as described in claim 1, characterized in that: In the dual-channel feature fusion analysis of step S4: the sensor mode noise extraction of channel one adopts the nonlocal mean denoising residual method, and the device-related noise components are separated under the guidance of the depth distribution map; The scene semantic relationship modeling of Channel 2 is based on the instance segmentation network to generate object-level semantic graphs, build a three-dimensional spatial relationship reasoning engine, and identify physical contradictions through shadow direction consistency detection and illumination reflection analysis; The attention weighting mechanism implements cross-channel feature saliency evaluation, specifically by dynamically allocating channel weights through a differentiable feature selector; the multi-directional sub-band energy distribution anomaly detection in step S5 includes: constructing a mutual information entropy measure between directional sub-bands and quantifying frequency domain correlation attenuation anomalies; A hierarchical anomaly response feature extractor is employed to capture frequency domain representations of forgery traces at multiple decomposition scales.

4. An image processing-based forensics system, used to implement the image processing-based forensics method as described in any one of claims 1-3, characterized in that, It includes a depth estimation module: containing a multi-scale encoder-decoder network, whose decoder embeds an edge-aware optimization layer for outputting a high-precision depth distribution map, the optimization layer being trained using a local gradient consistency loss function; Physical rule verification module: configured to perform focal length-object distance-depth mapping calculation, and generate a first-level tampering alarm signal when the difference exceeds the threshold by comparing the difference between the defocus blur gradient direction of the target area and the background area. Key region localization engine: includes: edge detection unit to identify depth-of-field transition boundaries, texture analysis unit to calculate local LBP entropy value, EXIF ​​verification unit to associate lens parameters with physical space logic, and filter key pixel regions; The dual-channel feature analysis module includes a first channel and a second channel. The first channel has a guided filter and a pattern noise extractor, while the second channel uses a scene semantic relationship analyzer driven by a graph convolutional network and an attention-weighted fusion unit to output a tampering suspicion score. The first-level alarm signal of the physical rule verification module directly outputs the tampering result; otherwise, it activates the second-level verification unit, which includes: a Steerable Pyramid multi-directional subband energy analyzer, a pre-trained SVM classifier, and a heatmap generator; the preprocessing interface is physically connected to the image sensor to acquire RAW format data and transmit it to the depth estimation module; the output priority of the physical rule verification module is higher than that of the dual-channel feature analysis module.

Images